From 189c6d42264cd735bb29a12499ea2f619fabac82 Mon Sep 17 00:00:00 2001 From: mtayfur Date: Mon, 27 Oct 2025 00:18:26 +0300 Subject: [PATCH 1/4] =?UTF-8?q?=F0=9F=94=A7=20(dev-check.sh,=20pyproject.t?= =?UTF-8?q?oml,=20requirements.txt):=20add=20development=20tooling=20and?= =?UTF-8?q?=20configuration?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Introduce a dev-check.sh script to automate code formatting and import sorting using Black and isort. Add a pyproject.toml file to configure Black and isort settings for consistent code style. Update requirements.txt to include Black and isort as development dependencies and remove version pinning for easier dependency management. These changes streamline the development workflow, enforce code style consistency, and make it easier for contributors to run formatting and import checks locally. --- dev-check.sh | 25 +++++++++++++++++++++++++ pyproject.toml | 24 ++++++++++++++++++++++++ requirements.txt | 10 ++++++---- 3 files changed, 55 insertions(+), 4 deletions(-) create mode 100755 dev-check.sh create mode 100644 pyproject.toml diff --git a/dev-check.sh b/dev-check.sh new file mode 100755 index 0000000..7241249 --- /dev/null +++ b/dev-check.sh @@ -0,0 +1,25 @@ +#!/usr/bin/env bash + +# Development tools script for openwebui-memory-system + +set -e + +if [ -f "./.venv/bin/python" ]; then + PYTHON="./.venv/bin/python" +elif command -v python3 &> /dev/null; then + PYTHON="python3" +elif command -v python &> /dev/null; then + PYTHON="python" +else + echo "Python 3 is not installed. Please install Python 3 to proceed." + exit 1 +fi + +echo "πŸ”§ Running development tools..." + +echo "🎨 Formatting with Black..." +$PYTHON -m black . +echo "πŸ“¦ Sorting imports with isort..." +$PYTHON -m isort . + +echo "βœ… All checks passed!" \ No newline at end of file diff --git a/pyproject.toml b/pyproject.toml new file mode 100644 index 0000000..9217c6d --- /dev/null +++ b/pyproject.toml @@ -0,0 +1,24 @@ +[tool.black] +line-length = 160 +target-version = ['py38'] +include = '\.pyi?$' +extend-exclude = ''' +/( + \.eggs + | \.git + | \.hg + | \.mypy_cache + | \.tox + | \.venv + | build + | dist +)/ +''' + +[tool.isort] +line_length = 160 +multi_line_output = 3 +include_trailing_comma = true +force_grid_wrap = 0 +use_parentheses = true +ensure_newline_before_comments = true \ No newline at end of file diff --git a/requirements.txt b/requirements.txt index e8f6881..910811c 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,5 +1,7 @@ -aiohttp>=3.12.15 -pydantic>=2.11.7 -numpy>=2.0.0 open-webui>=0.6.32 -tiktoken>=0.11.0 +aiohttp +pydantic +numpy +tiktoken +black +isort \ No newline at end of file From bb1bd012220d84710455a7afb29bb354b7a8de08 Mon Sep 17 00:00:00 2001 From: mtayfur Date: Mon, 27 Oct 2025 00:20:35 +0300 Subject: [PATCH 2/4] =?UTF-8?q?=E2=99=BB=EF=B8=8F=20(memory=5Fsystem.py):?= =?UTF-8?q?=20reformat=20code=20for=20consistency,=20readability,=20and=20?= =?UTF-8?q?maintainability?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - Reorder and group imports for clarity and PEP8 compliance. - Standardize string quoting and whitespace for consistency. - Refactor long function signatures and dictionary constructions for better readability. - Use double quotes for all string literals and dictionary keys. - Improve formatting of multiline statements and function calls. - Add or adjust line breaks to keep lines within recommended length. - Reformat class and method docstrings for clarity. - Use consistent indentation and spacing throughout the file. These changes improve code readability, maintainability, and consistency, making it easier for future contributors to understand and modify the codebase. No functional logic is changed. --- memory_system.py | 423 +++++++++++++++++++++++------------------------ 1 file changed, 205 insertions(+), 218 deletions(-) diff --git a/memory_system.py b/memory_system.py index 60d09ec..32adfe9 100644 --- a/memory_system.py +++ b/memory_system.py @@ -15,21 +15,22 @@ from enum import Enum from typing import Any, Callable, Dict, List, Optional, Tuple, Union import numpy as np -from pydantic import BaseModel, ConfigDict, Field, ValidationError as PydanticValidationError - -from open_webui.utils.chat import generate_chat_completion +from fastapi import Request from open_webui.models.users import Users from open_webui.routers.memories import Memories -from fastapi import Request +from open_webui.utils.chat import generate_chat_completion +from pydantic import BaseModel, ConfigDict, Field +from pydantic import ValidationError as PydanticValidationError logger = logging.getLogger(__name__) _SHARED_SKIP_DETECTOR_CACHE = {} _SHARED_SKIP_DETECTOR_CACHE_LOCK = asyncio.Lock() + class Constants: """Centralized configuration constants for the memory system.""" - + # Core System Limits MAX_MEMORY_CONTENT_CHARS = 500 # Character limit for LLM prompt memory content MAX_MEMORIES_PER_RETRIEVAL = 10 # Maximum memories returned per query @@ -37,31 +38,32 @@ class Constants: MIN_MESSAGE_CHARS = 10 # Minimum message length for validation DATABASE_OPERATION_TIMEOUT_SEC = 10 # Timeout for DB operations like user lookup LLM_CONSOLIDATION_TIMEOUT_SEC = 60.0 # Timeout for LLM consolidation operations - + # Cache System MAX_CACHE_ENTRIES_PER_TYPE = 500 # Maximum cache entries per cache type MAX_CONCURRENT_USER_CACHES = 50 # Maximum concurrent user cache instances CACHE_KEY_HASH_PREFIX_LENGTH = 10 # Hash prefix length for cache keys - + # Retrieval & Similarity - SEMANTIC_RETRIEVAL_THRESHOLD = 0.25 # Semantic similarity threshold for retrieval - RELAXED_SEMANTIC_THRESHOLD_MULTIPLIER = 0.8 # Multiplier for relaxed similarity threshold in secondary operations - EXTENDED_MAX_MEMORY_MULTIPLIER = 1.6 # Multiplier for expanding memory candidates in advanced operations - LLM_RERANKING_TRIGGER_MULTIPLIER = 0.8 # Multiplier for LLM reranking trigger threshold - + SEMANTIC_RETRIEVAL_THRESHOLD = 0.25 # Semantic similarity threshold for retrieval + RELAXED_SEMANTIC_THRESHOLD_MULTIPLIER = 0.8 # Multiplier for relaxed similarity threshold in secondary operations + EXTENDED_MAX_MEMORY_MULTIPLIER = 1.6 # Multiplier for expanding memory candidates in advanced operations + LLM_RERANKING_TRIGGER_MULTIPLIER = 0.8 # Multiplier for LLM reranking trigger threshold + # Skip Detection SKIP_CATEGORY_MARGIN = 0.5 # Margin above conversational similarity for skip category classification # Safety & Operations MAX_DELETE_OPERATIONS_RATIO = 0.6 # Maximum delete operations ratio for safety MIN_OPS_FOR_DELETE_RATIO_CHECK = 6 # Minimum operations to apply ratio check - + # Content Display - CONTENT_PREVIEW_LENGTH = 80 # Maximum length for content preview display - + CONTENT_PREVIEW_LENGTH = 80 # Maximum length for content preview display + # Default Models DEFAULT_LLM_MODEL = "google/gemini-2.5-flash-lite" + class Prompts: """Container for all LLM prompts used in the memory system.""" @@ -181,6 +183,7 @@ Return: {{"ids": []}} Explanation: Query seeks general technical explanation without personal context. Job and family information don't affect how quantum computing concepts should be explained. """ + class Models: """Container for all Pydantic models used in the memory system.""" @@ -203,7 +206,7 @@ class Models: class MemoryOperation(StrictModel): """Pydantic model for memory operations with validation.""" - operation: 'Models.MemoryOperationType' = Field(description="Type of memory operation to perform") + operation: "Models.MemoryOperationType" = Field(description="Type of memory operation to perform") content: str = Field(description="Memory content (required for CREATE/UPDATE, empty for DELETE)") id: str = Field(description="Memory ID (empty for CREATE, required for UPDATE/DELETE)") @@ -221,7 +224,7 @@ class Models: class ConsolidationResponse(BaseModel): """Pydantic model for memory consolidation LLM response - object containing array of memory operations.""" - ops: List['Models.MemoryOperation'] = Field(default_factory=list, description="List of memory operations to execute") + ops: List["Models.MemoryOperation"] = Field(default_factory=list, description="List of memory operations to execute") class MemoryRerankingResponse(BaseModel): """Pydantic model for memory reranking LLM response - object containing array of memory IDs.""" @@ -442,52 +445,42 @@ class SkipDetector: self.embedding_function = embedding_function self._reference_embeddings = None self._initialize_reference_embeddings() - + def _initialize_reference_embeddings(self) -> None: """Compute and cache embeddings for category descriptions.""" try: - technical_embeddings = self.embedding_function( - self.TECHNICAL_CATEGORY_DESCRIPTIONS - ) - - instruction_embeddings = self.embedding_function( - self.INSTRUCTION_CATEGORY_DESCRIPTIONS - ) - - pure_math_embeddings = self.embedding_function( - self.PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS - ) - - translation_embeddings = self.embedding_function( - self.EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS - ) - - grammar_embeddings = self.embedding_function( - self.GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS - ) - - conversational_embeddings = self.embedding_function( - self.CONVERSATIONAL_CATEGORY_DESCRIPTIONS - ) - + technical_embeddings = self.embedding_function(self.TECHNICAL_CATEGORY_DESCRIPTIONS) + + instruction_embeddings = self.embedding_function(self.INSTRUCTION_CATEGORY_DESCRIPTIONS) + + pure_math_embeddings = self.embedding_function(self.PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS) + + translation_embeddings = self.embedding_function(self.EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS) + + grammar_embeddings = self.embedding_function(self.GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS) + + conversational_embeddings = self.embedding_function(self.CONVERSATIONAL_CATEGORY_DESCRIPTIONS) + self._reference_embeddings = { - 'technical': np.array(technical_embeddings), - 'instruction': np.array(instruction_embeddings), - 'pure_math': np.array(pure_math_embeddings), - 'translation': np.array(translation_embeddings), - 'grammar': np.array(grammar_embeddings), - 'conversational': np.array(conversational_embeddings), + "technical": np.array(technical_embeddings), + "instruction": np.array(instruction_embeddings), + "pure_math": np.array(pure_math_embeddings), + "translation": np.array(translation_embeddings), + "grammar": np.array(grammar_embeddings), + "conversational": np.array(conversational_embeddings), } - + total_skip_categories = ( - len(self.TECHNICAL_CATEGORY_DESCRIPTIONS) + - len(self.INSTRUCTION_CATEGORY_DESCRIPTIONS) + - len(self.PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS) + - len(self.EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS) + - len(self.GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS) + len(self.TECHNICAL_CATEGORY_DESCRIPTIONS) + + len(self.INSTRUCTION_CATEGORY_DESCRIPTIONS) + + len(self.PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS) + + len(self.EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS) + + len(self.GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS) + ) + + logger.info( + f"SkipDetector initialized with {total_skip_categories} skip categories and {len(self.CONVERSATIONAL_CATEGORY_DESCRIPTIONS)} personal categories" ) - - logger.info(f"SkipDetector initialized with {total_skip_categories} skip categories and {len(self.CONVERSATIONAL_CATEGORY_DESCRIPTIONS)} personal categories") except Exception as e: logger.error(f"Failed to initialize SkipDetector reference embeddings: {e}") self._reference_embeddings = None @@ -504,108 +497,107 @@ class SkipDetector: def _fast_path_skip_detection(self, message: str) -> Optional[str]: """Language-agnostic structural pattern detection with high confidence and low false positive rate.""" msg_len = len(message) - + # Pattern 1: Multiple URLs (5+ full URLs indicates link lists or technical references) - url_pattern_count = message.count('http://') + message.count('https://') + url_pattern_count = message.count("http://") + message.count("https://") if url_pattern_count >= 5: return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 2: Long unbroken alphanumeric strings (tokens, hashes, base64) words = message.split() for word in words: cleaned = word.strip('.,;:!?()[]{}"\'"') - if len(cleaned) > 80 and cleaned.replace('-', '').replace('_', '').isalnum(): + if len(cleaned) > 80 and cleaned.replace("-", "").replace("_", "").isalnum(): return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 3: Markdown/text separators (repeated ---, ===, ___, ***) - separator_patterns = ['---', '===', '___', '***'] + separator_patterns = ["---", "===", "___", "***"] for pattern in separator_patterns: if message.count(pattern) >= 2: return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 4: Command-line patterns with context-aware detection - lines_stripped = [line.strip() for line in message.split('\n') if line.strip()] + lines_stripped = [line.strip() for line in message.split("\n") if line.strip()] if lines_stripped: actual_command_lines = 0 for line in lines_stripped: - if line.startswith('$ ') and len(line) > 2: + if line.startswith("$ ") and len(line) > 2: parts = line[2:].split() if parts and parts[0].isalnum(): actual_command_lines += 1 - elif '$ ' in line: - dollar_index = line.find('$ ') - if dollar_index > 0 and line[dollar_index-1] in (' ', ':', '\t'): - parts = line[dollar_index+2:].split() - if parts and len(parts[0]) > 0 and (parts[0].isalnum() or parts[0] in ['curl', 'wget', 'git', 'npm', 'pip', 'docker']): + elif "$ " in line: + dollar_index = line.find("$ ") + if dollar_index > 0 and line[dollar_index - 1] in (" ", ":", "\t"): + parts = line[dollar_index + 2 :].split() + if parts and len(parts[0]) > 0 and (parts[0].isalnum() or parts[0] in ["curl", "wget", "git", "npm", "pip", "docker"]): actual_command_lines += 1 - elif line.startswith('# ') and len(line) > 2: + elif line.startswith("# ") and len(line) > 2: rest = line[2:].strip() - if rest and not rest[0].isupper() and ' ' in rest: + if rest and not rest[0].isupper() and " " in rest: actual_command_lines += 1 - elif line.startswith('> ') and len(line) > 2: + elif line.startswith("> ") and len(line) > 2: pass - - if actual_command_lines >= 1 and any(c in message for c in ['http://', 'https://', ' | ']): + + if actual_command_lines >= 1 and any(c in message for c in ["http://", "https://", " | "]): return self.SkipReason.SKIP_TECHNICAL.value if actual_command_lines >= 3: return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 5: High path/URL density (dots and slashes suggesting file paths or URLs) if msg_len > 30: - slash_count = message.count('/') + message.count('\\') - dot_count = message.count('.') + slash_count = message.count("/") + message.count("\\") + dot_count = message.count(".") path_chars = slash_count + dot_count if path_chars > 10 and (path_chars / msg_len) > 0.15: return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 6: Markup character density (structured data) - markup_chars = sum(message.count(c) for c in '{}[]<>') + markup_chars = sum(message.count(c) for c in "{}[]<>") if markup_chars >= 6: if markup_chars / msg_len > 0.10: return self.SkipReason.SKIP_TECHNICAL.value - curly_count = message.count('{') + message.count('}') + curly_count = message.count("{") + message.count("}") if curly_count >= 10: return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 7: Structured nested content with colons (key: value patterns) - line_count = message.count('\n') + line_count = message.count("\n") if line_count >= 8: - lines = message.split('\n') + lines = message.split("\n") non_empty_lines = [line for line in lines if line.strip()] if non_empty_lines: - colon_lines = sum(1 for line in non_empty_lines if ':' in line and not line.strip().startswith('#')) - indented_lines = sum(1 for line in non_empty_lines if line.startswith((' ', '\t'))) - - if (colon_lines / len(non_empty_lines) > 0.4 and - indented_lines / len(non_empty_lines) > 0.5): + colon_lines = sum(1 for line in non_empty_lines if ":" in line and not line.strip().startswith("#")) + indented_lines = sum(1 for line in non_empty_lines if line.startswith((" ", "\t"))) + + if colon_lines / len(non_empty_lines) > 0.4 and indented_lines / len(non_empty_lines) > 0.5: return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 8: Highly structured multi-line content (require markup chars for technical confidence) if line_count > 15: - lines = message.split('\n') + lines = message.split("\n") non_empty_lines = [line for line in lines if line.strip()] if non_empty_lines: - markup_in_lines = sum(1 for line in non_empty_lines if any(c in line for c in '{}[]<>')) - structured_lines = sum(1 for line in non_empty_lines if line.startswith((' ', '\t'))) - + markup_in_lines = sum(1 for line in non_empty_lines if any(c in line for c in "{}[]<>")) + structured_lines = sum(1 for line in non_empty_lines if line.startswith((" ", "\t"))) + if markup_in_lines / len(non_empty_lines) > 0.3: return self.SkipReason.SKIP_TECHNICAL.value elif structured_lines / len(non_empty_lines) > 0.6: - technical_keywords = ['function', 'class', 'import', 'return', 'const', 'var', 'let', 'def'] + technical_keywords = ["function", "class", "import", "return", "const", "var", "let", "def"] if any(keyword in message.lower() for keyword in technical_keywords): return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 9: Code-like indentation pattern (require code indicators to avoid false positives from bullet lists) if line_count >= 3: - lines = message.split('\n') + lines = message.split("\n") non_empty_lines = [line for line in lines if line.strip()] if non_empty_lines: - indented_lines = sum(1 for line in non_empty_lines if line[0] in (' ', '\t')) + indented_lines = sum(1 for line in non_empty_lines if line[0] in (" ", "\t")) if indented_lines / len(non_empty_lines) > 0.5: - code_indicators = ['def ', 'class ', 'function ', 'return ', 'import ', 'const ', 'let ', 'var ', 'public ', 'private '] + code_indicators = ["def ", "class ", "function ", "return ", "import ", "const ", "let ", "var ", "public ", "private "] if any(indicator in message.lower() for indicator in code_indicators): return self.SkipReason.SKIP_TECHNICAL.value - + # Pattern 10: Very high special character ratio (encoded data, technical output) if msg_len > 50: special_chars = sum(1 for c in message if not c.isalnum() and not c.isspace()) @@ -614,10 +606,10 @@ class SkipDetector: alphanumeric = sum(1 for c in message if c.isalnum()) if alphanumeric / msg_len < 0.50: return self.SkipReason.SKIP_TECHNICAL.value - + return None - def detect_skip_reason(self, message: str, max_message_chars: int, memory_system: 'Filter') -> Optional[str]: + def detect_skip_reason(self, message: str, max_message_chars: int, memory_system: "Filter") -> Optional[str]: """ Detect if a message should be skipped using two-stage detection: 1. Fast-path structural patterns (~95% confidence) @@ -628,53 +620,49 @@ class SkipDetector: size_issue = self.validate_message_size(message, max_message_chars) if size_issue: return size_issue - + fast_skip = self._fast_path_skip_detection(message) if fast_skip: logger.info(f"Fast-path skip: {fast_skip}") return fast_skip - + if self._reference_embeddings is None: logger.warning("SkipDetector reference embeddings not initialized, allowing message through") return None - + try: message_embedding = np.array(self.embedding_function([message.strip()])[0]) - - conversational_similarities = np.dot( - message_embedding, - self._reference_embeddings['conversational'].T - ) + + conversational_similarities = np.dot(message_embedding, self._reference_embeddings["conversational"].T) max_conversational_similarity = float(conversational_similarities.max()) - + skip_categories = [ - ('instruction', self.SkipReason.SKIP_INSTRUCTION, self.INSTRUCTION_CATEGORY_DESCRIPTIONS), - ('translation', self.SkipReason.SKIP_TRANSLATION, self.EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS), - ('grammar', self.SkipReason.SKIP_GRAMMAR_PROOFREAD, self.GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS), - ('technical', self.SkipReason.SKIP_TECHNICAL, self.TECHNICAL_CATEGORY_DESCRIPTIONS), - ('pure_math', self.SkipReason.SKIP_PURE_MATH, self.PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS), + ("instruction", self.SkipReason.SKIP_INSTRUCTION, self.INSTRUCTION_CATEGORY_DESCRIPTIONS), + ("translation", self.SkipReason.SKIP_TRANSLATION, self.EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS), + ("grammar", self.SkipReason.SKIP_GRAMMAR_PROOFREAD, self.GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS), + ("technical", self.SkipReason.SKIP_TECHNICAL, self.TECHNICAL_CATEGORY_DESCRIPTIONS), + ("pure_math", self.SkipReason.SKIP_PURE_MATH, self.PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS), ] - + qualifying_categories = [] margin_threshold = max_conversational_similarity + Constants.SKIP_CATEGORY_MARGIN - + for cat_key, skip_reason, descriptions in skip_categories: - similarities = np.dot( - message_embedding, - self._reference_embeddings[cat_key].T - ) + similarities = np.dot(message_embedding, self._reference_embeddings[cat_key].T) max_similarity = float(similarities.max()) - + if max_similarity > margin_threshold: qualifying_categories.append((max_similarity, cat_key, skip_reason)) - + if qualifying_categories: highest_similarity, highest_cat_key, highest_skip_reason = max(qualifying_categories, key=lambda x: x[0]) - logger.info(f"🚫 Skipping message: {highest_skip_reason.value} (sim {highest_similarity:.3f} > conv {max_conversational_similarity:.3f} + {Constants.SKIP_CATEGORY_MARGIN:.3f})") + logger.info( + f"🚫 Skipping message: {highest_skip_reason.value} (sim {highest_similarity:.3f} > conv {max_conversational_similarity:.3f} + {Constants.SKIP_CATEGORY_MARGIN:.3f})" + ) return highest_skip_reason.value - + return None - + except Exception as e: logger.error(f"Error in semantic skip detection: {e}") return None @@ -692,7 +680,7 @@ class LLMRerankingService: llm_trigger_threshold = int(self.memory_system.valves.max_memories_returned * self.memory_system.valves.llm_reranking_trigger_multiplier) if len(memories) > llm_trigger_threshold: - return True, f"{len(memories)} candidate memories exceed {llm_trigger_threshold} threshold" + return True, f"{len(memories)} candidate memories exceed {llm_trigger_threshold} threshold" return False, f"{len(memories)} candidate memories within threshold of {llm_trigger_threshold}" @@ -717,33 +705,29 @@ CANDIDATE MEMORIES: selected_memories.append(memory) logger.info(f"🧠 LLM selected {len(selected_memories)} out of {len(candidate_memories)} candidates") - + return selected_memories except Exception as e: - logger.warning(f"πŸ€– LLM reranking failed during memory relevance analysis: {str(e)}") + logger.warning(f"πŸ€– LLM reranking failed during memory relevance analysis: {str(e)}") return candidate_memories - async def rerank_memories( - self, user_message: str, candidate_memories: List[Dict], emitter: Optional[Callable] = None - ) -> Tuple[List[Dict], Dict[str, Any]]: + async def rerank_memories(self, user_message: str, candidate_memories: List[Dict], emitter: Optional[Callable] = None) -> Tuple[List[Dict], Dict[str, Any]]: start_time = time.time() max_injection = self.memory_system.valves.max_memories_returned should_use_llm, decision_reason = self._should_use_llm_reranking(candidate_memories) - analysis_info = {"llm_decision": should_use_llm, "decision_reason": decision_reason, "candidate_count": len(candidate_memories)} + analysis_info = {"llm_decision": should_use_llm, "decision_reason": decision_reason, "candidate_count": len(candidate_memories)} if should_use_llm: extended_count = int(self.memory_system.valves.max_memories_returned * Constants.EXTENDED_MAX_MEMORY_MULTIPLIER) llm_candidates = candidate_memories[:extended_count] - await self.memory_system._emit_status( - emitter, f"πŸ€– LLM Analyzing {len(llm_candidates)} Memories for Relevance", done=False - ) + await self.memory_system._emit_status(emitter, f"πŸ€– LLM Analyzing {len(llm_candidates)} Memories for Relevance", done=False) logger.info(f"Using LLM reranking: {decision_reason}") selected_memories = await self._llm_select_memories(user_message, llm_candidates, max_injection, emitter) - + if not selected_memories: logger.info("πŸ“­ No relevant memories after LLM analysis") await self.memory_system._emit_status(emitter, f"πŸ“­ No Relevant Memories After LLM Analysis", done=True) @@ -751,7 +735,7 @@ CANDIDATE MEMORIES: else: logger.info(f"Skipping LLM reranking: {decision_reason}") selected_memories = candidate_memories[:max_injection] - + duration = time.time() - start_time duration_text = f" in {duration:.2f}s" if duration >= 0.01 else "" retrieval_method = "LLM" if should_use_llm else "Semantic" @@ -769,10 +753,10 @@ class LLMConsolidationService: """Filter consolidation candidates by threshold and return candidates with threshold info.""" consolidation_threshold = self.memory_system._get_retrieval_threshold(is_consolidation=True) candidates = [mem for mem in similarities if mem["relevance"] >= consolidation_threshold] - + max_consolidation_memories = int(self.memory_system.valves.max_memories_returned * Constants.EXTENDED_MAX_MEMORY_MULTIPLIER) candidates = candidates[:max_consolidation_memories] - + threshold_info = f"{consolidation_threshold:.3f} (max: {max_consolidation_memories})" return candidates, threshold_info @@ -812,7 +796,7 @@ class LLMConsolidationService: candidates, threshold_info = self._filter_consolidation_candidates(all_similarities) else: candidates = [] - threshold_info = 'N/A' + threshold_info = "N/A" logger.info(f"🎯 Found {len(candidates)} candidate memories for consolidation (threshold: {threshold_info})") @@ -820,7 +804,9 @@ class LLMConsolidationService: return candidates - async def generate_consolidation_plan(self, user_message: str, candidate_memories: List[Dict[str, Any]], emitter: Optional[Callable] = None) -> List[Dict[str, Any]]: + async def generate_consolidation_plan( + self, user_message: str, candidate_memories: List[Dict[str, Any]], emitter: Optional[Callable] = None + ) -> List[Dict[str, Any]]: """Generate consolidation plan using LLM with clear system/user prompt separation.""" if candidate_memories: memory_lines = self.memory_system._format_memories_for_llm(candidate_memories) @@ -925,7 +911,7 @@ class LLMConsolidationService: results = await asyncio.gather(*batch_tasks, return_exceptions=True) for idx, result in enumerate(results): operation = ops[idx] - + if isinstance(result, Exception): failed_count += 1 await self.memory_system._emit_status(emitter, f"❌ Failed {operation_type}", done=False) @@ -982,21 +968,21 @@ class LLMConsolidationService: if operations: created_count, updated_count, deleted_count, failed_count = await self.execute_memory_operations(operations, user_id, emitter) - + duration = time.time() - start_time logger.info(f"πŸ’Ύ Memory Consolidation Complete In {duration:.2f}s") - + total_operations = created_count + updated_count + deleted_count if total_operations > 0 or failed_count > 0: await self.memory_system._emit_status(emitter, f"πŸ’Ύ Memory Consolidation Complete in {duration:.2f}s", done=False) - + operation_details = self.memory_system._build_operation_details(created_count, updated_count, deleted_count) memory_word = "Memory" if total_operations == 1 else "Memories" operations_summary = f"{', '.join(operation_details)} {memory_word}" - + if failed_count > 0: operations_summary += f" (❌ {failed_count} Failed)" - + await self.memory_system._emit_status(emitter, operations_summary, done=True) except Exception as e: @@ -1016,20 +1002,27 @@ class Filter: """Configuration valves for the Memory System.""" model: str = Field(default=Constants.DEFAULT_LLM_MODEL, description="Model name for LLM operations") - + max_message_chars: int = Field(default=Constants.MAX_MESSAGE_CHARS, description="Maximum user message length before skipping memory operations") max_memories_returned: int = Field(default=Constants.MAX_MEMORIES_PER_RETRIEVAL, description="Maximum number of memories to return in context") - - semantic_retrieval_threshold: float = Field(default=Constants.SEMANTIC_RETRIEVAL_THRESHOLD, description="Minimum similarity threshold for memory retrieval") - relaxed_semantic_threshold_multiplier: float = Field(default=Constants.RELAXED_SEMANTIC_THRESHOLD_MULTIPLIER, description="Adjusts similarity threshold for memory consolidation (lower = more candidates)") - + + semantic_retrieval_threshold: float = Field( + default=Constants.SEMANTIC_RETRIEVAL_THRESHOLD, description="Minimum similarity threshold for memory retrieval" + ) + relaxed_semantic_threshold_multiplier: float = Field( + default=Constants.RELAXED_SEMANTIC_THRESHOLD_MULTIPLIER, + description="Adjusts similarity threshold for memory consolidation (lower = more candidates)", + ) + enable_llm_reranking: bool = Field(default=True, description="Enable LLM-based memory reranking for improved contextual selection") - llm_reranking_trigger_multiplier: float = Field(default=Constants.LLM_RERANKING_TRIGGER_MULTIPLIER, description="Controls when LLM reranking activates (lower = more aggressive)") + llm_reranking_trigger_multiplier: float = Field( + default=Constants.LLM_RERANKING_TRIGGER_MULTIPLIER, description="Controls when LLM reranking activates (lower = more aggressive)" + ) def __init__(self): """Initialize the Memory System filter with production validation.""" global _SHARED_SKIP_DETECTOR_CACHE - + self.valves = self.Valves() self._validate_system_configuration() @@ -1043,8 +1036,13 @@ class Filter: self._llm_reranking_service = LLMRerankingService(self) self._llm_consolidation_service = LLMConsolidationService(self) - async def _set_pipeline_context(self, __event_emitter__: Optional[Callable] = None, __user__: Optional[Dict[str, Any]] = None, - __model__: Optional[str] = None, __request__: Optional[Request] = None) -> None: + async def _set_pipeline_context( + self, + __event_emitter__: Optional[Callable] = None, + __user__: Optional[Dict[str, Any]] = None, + __model__: Optional[str] = None, + __request__: Optional[Request] = None, + ) -> None: """Set pipeline context parameters to avoid duplication in inlet/outlet methods.""" if __event_emitter__: self.__current_event_emitter__ = __event_emitter__ @@ -1054,17 +1052,17 @@ class Filter: self.__model__ = __model__ if __request__: self.__request__ = __request__ - - if self._embedding_function is None and hasattr(__request__.app.state, 'EMBEDDING_FUNCTION'): + + if self._embedding_function is None and hasattr(__request__.app.state, "EMBEDDING_FUNCTION"): self._embedding_function = __request__.app.state.EMBEDDING_FUNCTION logger.info(f"βœ… Using OpenWebUI's embedding function") - + if self._skip_detector is None: global _SHARED_SKIP_DETECTOR_CACHE, _SHARED_SKIP_DETECTOR_CACHE_LOCK - embedding_engine = getattr(__request__.app.state.config, 'RAG_EMBEDDING_ENGINE', '') - embedding_model = getattr(__request__.app.state.config, 'RAG_EMBEDDING_MODEL', '') + embedding_engine = getattr(__request__.app.state.config, "RAG_EMBEDDING_ENGINE", "") + embedding_model = getattr(__request__.app.state.config, "RAG_EMBEDDING_MODEL", "") cache_key = f"{embedding_engine}:{embedding_model}" - + async with _SHARED_SKIP_DETECTOR_CACHE_LOCK: if cache_key in _SHARED_SKIP_DETECTOR_CACHE: logger.info(f"♻️ Reusing cached skip detector: {cache_key}") @@ -1072,6 +1070,7 @@ class Filter: else: logger.info(f"πŸ€– Initializing skip detector with OpenWebUI embeddings: {cache_key}") embedding_fn = self._embedding_function + def embedding_wrapper(texts: Union[str, List[str]]) -> Union[np.ndarray, List[np.ndarray]]: result = embedding_fn(texts, prefix=None, user=None) if isinstance(result, list): @@ -1079,12 +1078,11 @@ class Filter: return [np.array(emb, dtype=np.float16) for emb in result] return np.array(result, dtype=np.float16) return np.array(result, dtype=np.float16) - + self._skip_detector = SkipDetector(embedding_wrapper) _SHARED_SKIP_DETECTOR_CACHE[cache_key] = self._skip_detector logger.info(f"βœ… Skip detector initialized and cached") - def _truncate_content(self, content: str, max_length: Optional[int] = None) -> str: """Truncate content with ellipsis if needed.""" if max_length is None: @@ -1148,7 +1146,7 @@ class Filter: """Unified embedding generation for single text or batch with optimized caching using OpenWebUI's embedding function.""" if self._embedding_function is None: raise RuntimeError("πŸ€– Embedding function not initialized. Ensure pipeline context is set.") - + is_single = isinstance(texts, str) text_list = [texts] if is_single else texts @@ -1181,17 +1179,11 @@ class Filter: uncached_hashes.append(text_hash) if uncached_texts: - user = await asyncio.to_thread(Users.get_user_by_id, user_id) if hasattr(self, '__user__') else None - + user = await asyncio.to_thread(Users.get_user_by_id, user_id) if hasattr(self, "__user__") else None + loop = asyncio.get_event_loop() - raw_embeddings = await loop.run_in_executor( - None, - self._embedding_function, - uncached_texts, - None, - user - ) - + raw_embeddings = await loop.run_in_executor(None, self._embedding_function, uncached_texts, None, user) + if isinstance(raw_embeddings, list) and len(raw_embeddings) > 0: if isinstance(raw_embeddings[0], list): new_embeddings = [self._normalize_embedding(emb) for emb in raw_embeddings] @@ -1211,15 +1203,13 @@ class Filter: return result_embeddings[0] else: valid_count = sum(1 for emb in result_embeddings if emb is not None) - logger.info( - f"πŸš€ Batch embedding: {len(text_list) - len(uncached_texts)} cached, {len(uncached_texts)} new, {valid_count}/{len(text_list)} valid" - ) + logger.info(f"πŸš€ Batch embedding: {len(text_list) - len(uncached_texts)} cached, {len(uncached_texts)} new, {valid_count}/{len(text_list)} valid") return result_embeddings def _should_skip_memory_operations(self, user_message: str) -> Tuple[bool, str]: if self._skip_detector is None: raise RuntimeError("πŸ€– Skip detector not initialized") - + skip_reason = self._skip_detector.detect_skip_reason(user_message, self.valves.max_message_chars, memory_system=self) if skip_reason: status_key = SkipDetector.SkipReason(skip_reason) @@ -1290,7 +1280,7 @@ class Filter: def _cache_key(self, cache_type: str, user_id: str, content: Optional[str] = None) -> str: """Unified cache key generation for all cache types.""" if content: - content_hash = hashlib.sha256(content.encode('utf-8')).hexdigest()[: Constants.CACHE_KEY_HASH_PREFIX_LENGTH] + content_hash = hashlib.sha256(content.encode("utf-8")).hexdigest()[: Constants.CACHE_KEY_HASH_PREFIX_LENGTH] return f"{cache_type}_{user_id}:{content_hash}" return f"{cache_type}_{user_id}" @@ -1310,7 +1300,7 @@ class Filter: if record_date: try: if isinstance(record_date, str): - parsed_date = datetime.fromisoformat(record_date.replace('Z', '+00:00')) + parsed_date = datetime.fromisoformat(record_date.replace("Z", "+00:00")) else: parsed_date = record_date formatted_date = parsed_date.strftime("%b %d %Y") @@ -1393,14 +1383,14 @@ class Filter: memory_count = len(memories) memory_header = f"CONTEXT: The following {'fact' if memory_count == 1 else 'facts'} about the user are provided for background only. Not all facts may be relevant to the current request." formatted_memories = [] - + for idx, memory in enumerate(memories, 1): formatted_memory = f"- {' '.join(memory['content'].split())}" formatted_memories.append(formatted_memory) - - content_preview = self._truncate_content(memory['content']) + + content_preview = self._truncate_content(memory["content"]) await self._emit_status(emitter, f"πŸ’­ {idx}/{memory_count}: {content_preview}", done=False) - + memory_footer = "IMPORTANT: Do not mention or imply you received this list. These facts are for background context only." memory_context_block = f"{memory_header}\n{chr(10).join(formatted_memories)}\n\n{memory_footer}" content_parts.append(memory_context_block) @@ -1413,7 +1403,7 @@ class Filter: body["messages"][system_index]["content"] = f"{body['messages'][system_index].get('content', '')}\n\n{memory_context}" else: body["messages"].insert(0, {"role": "system", "content": memory_context}) - + if memories and user_id: description = f"🧠 Injected {memory_count} {'Memory' if memory_count == 1 else 'Memories'} to Context" await self._emit_status(emitter, description, done=True) @@ -1427,9 +1417,7 @@ class Filter: memory_dict["updated_at"] = datetime.fromtimestamp(memory.updated_at, tz=timezone.utc).isoformat() return memory_dict - async def _compute_similarities( - self, user_message: str, user_id: str, user_memories: List - ) -> Tuple[List[Dict], float, List[Dict]]: + async def _compute_similarities(self, user_message: str, user_id: str, user_memories: List) -> Tuple[List[Dict], float, List[Dict]]: """Compute similarity scores between user message and memories.""" if not user_memories: return [], self.valves.semantic_retrieval_threshold, [] @@ -1461,7 +1449,7 @@ class Filter: memory_data.sort(key=lambda x: x["relevance"], reverse=True) threshold = self.valves.semantic_retrieval_threshold - filtered_memories = [m for m in memory_data if m["relevance"] >= threshold] + filtered_memories = [m for m in memory_data if m["relevance"] >= threshold] return filtered_memories, threshold, memory_data async def inlet( @@ -1536,9 +1524,7 @@ class Filter: cache_key = self._cache_key(self._cache_manager.RETRIEVAL_CACHE, user_id, user_message) cached_similarities = await self._cache_manager.get(user_id, self._cache_manager.RETRIEVAL_CACHE, cache_key) - task = asyncio.create_task( - self._llm_consolidation_service.run_consolidation_pipeline(user_message, user_id, __event_emitter__, cached_similarities) - ) + task = asyncio.create_task(self._llm_consolidation_service.run_consolidation_pipeline(user_message, user_id, __event_emitter__, cached_similarities)) self._background_tasks.add(task) def safe_cleanup(t: asyncio.Task) -> None: @@ -1582,11 +1568,7 @@ class Filter: await self._cache_manager.put(user_id, self._cache_manager.MEMORY_CACHE, memory_cache_key, user_memories) - memory_contents = [ - memory.content - for memory in user_memories - if memory.content and len(memory.content.strip()) >= Constants.MIN_MESSAGE_CHARS - ] + memory_contents = [memory.content for memory in user_memories if memory.content and len(memory.content.strip()) >= Constants.MIN_MESSAGE_CHARS] if memory_contents: await self._generate_embeddings(memory_contents, user_id) @@ -1615,7 +1597,7 @@ class Filter: if not id_stripped: logger.warning(f"⚠️ Skipping UPDATE operation: empty ID") return Models.OperationResult.SKIPPED_EMPTY_ID.value - + content_stripped = operation.content.strip() if not content_stripped: logger.warning(f"⚠️ Skipping UPDATE operation for {id_stripped}: empty content") @@ -1649,18 +1631,18 @@ class Filter: """Remove $ref references and ensure required fields for Azure OpenAI.""" if not isinstance(schema, dict): return schema - - if '$ref' in schema: - ref_path = schema['$ref'] - if ref_path.startswith('#/$defs/'): - def_name = ref_path.split('/')[-1] + + if "$ref" in schema: + ref_path = schema["$ref"] + if ref_path.startswith("#/$defs/"): + def_name = ref_path.split("/")[-1] if schema_defs and def_name in schema_defs: return self._remove_refs_from_schema(schema_defs[def_name].copy(), schema_defs) - return {'type': 'object'} - + return {"type": "object"} + result = {} for key, value in schema.items(): - if key == '$defs': + if key == "$defs": continue elif isinstance(value, dict): result[key] = self._remove_refs_from_schema(value, schema_defs) @@ -1668,10 +1650,10 @@ class Filter: result[key] = [self._remove_refs_from_schema(item, schema_defs) if isinstance(item, dict) else item for item in value] else: result[key] = value - - if result.get('type') == 'object' and 'properties' in result: - result['required'] = list(result['properties'].keys()) - + + if result.get("type") == "object" and "properties" in result: + result["required"] = list(result["properties"].keys()) + return result async def _query_llm(self, system_prompt: str, user_prompt: str, response_model: Optional[BaseModel] = None) -> Union[str, BaseModel]: @@ -1685,16 +1667,16 @@ class Filter: form_data = { "model": model_to_use, - "messages": [{"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}], + "messages": [{"role": "system", "content": system_prompt}, {"role": "user", "content": user_prompt}], "max_tokens": 4096, "stream": False, } if response_model: raw_schema = response_model.model_json_schema() - schema_defs = raw_schema.get('$defs', {}) + schema_defs = raw_schema.get("$defs", {}) schema = self._remove_refs_from_schema(raw_schema, schema_defs) - schema['type'] = 'object' + schema["type"] = "object" form_data["response_format"] = {"type": "json_schema", "json_schema": {"name": response_model.__name__, "strict": True, "schema": schema}} try: @@ -1718,7 +1700,12 @@ class Filter: if isinstance(response_data, dict) and "choices" in response_data and isinstance(response_data["choices"], list) and len(response_data["choices"]) > 0: first_choice = response_data["choices"][0] - if isinstance(first_choice, dict) and "message" in first_choice and isinstance(first_choice["message"], dict) and "content" in first_choice["message"]: + if ( + isinstance(first_choice, dict) + and "message" in first_choice + and isinstance(first_choice["message"], dict) + and "content" in first_choice["message"] + ): content = first_choice["message"]["content"] else: raise ValueError("πŸ€– Invalid response structure: missing content in message") @@ -1726,7 +1713,7 @@ class Filter: raise ValueError(f"πŸ€– Unexpected LLM response format: {response_data}") if response_model: - try: + try: parsed_data = json.loads(content) return response_model.model_validate(parsed_data) except json.JSONDecodeError as e: From 3f9b4c6d48cfeeeaf20205444785e371cf7db83c Mon Sep 17 00:00:00 2001 From: mtayfur Date: Sun, 26 Oct 2025 23:31:37 +0300 Subject: [PATCH 3/4] =?UTF-8?q?=E2=99=BB=EF=B8=8F=20(memory=5Fsystem):=20r?= =?UTF-8?q?efactor=20skip=20detection=20and=20add=20semantic=20deduplicati?= =?UTF-8?q?on?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - Unify skip detection to a binary classifier (personal vs non-personal) for improved maintainability and clarity. Remove multiple technical/ instruction/translation/etc. categories and consolidate into NON_PERSONAL and PERSONAL. - Adjust skip detection margin for more precise classification. - Add semantic deduplication for memory operations using embedding similarity, preventing duplicate memory creation and updates. - Normalize and validate embedding dimensions for robustness. - Add per-user async locks to prevent race conditions during memory consolidation. - Refactor requirements.txt to remove version pinning for easier dependency management. - Improve logging and error handling for embedding and deduplication operations. These changes improve the reliability and accuracy of memory classification and deduplication, reduce false positives in skip detection, and prevent duplicate or conflicting memory operations in concurrent environments. Dependency management is simplified for compatibility. --- .gitignore | 2 +- memory_system.py | 173 ++++++++++++++++++++++++++++++++++++++--------- requirements.txt | 2 +- 3 files changed, 143 insertions(+), 34 deletions(-) diff --git a/.gitignore b/.gitignore index f2d0f25..3ec1fec 100644 --- a/.gitignore +++ b/.gitignore @@ -1,4 +1,4 @@ __pycache__/ +.github/instructions/* .venv/ -**AGENTS.md tests/ \ No newline at end of file diff --git a/memory_system.py b/memory_system.py index 32adfe9..394a634 100644 --- a/memory_system.py +++ b/memory_system.py @@ -1,6 +1,9 @@ """ title: Memory System +description: A semantic memory management system for Open WebUI that consolidates, deduplicates, and retrieves personalized user memories using LLM operations. version: 1.0.0 +authors: https://github.com/mtayfur +license: Apache-2.0 """ import asyncio @@ -51,7 +54,7 @@ class Constants: LLM_RERANKING_TRIGGER_MULTIPLIER = 0.8 # Multiplier for LLM reranking trigger threshold # Skip Detection - SKIP_CATEGORY_MARGIN = 0.5 # Margin above conversational similarity for skip category classification + SKIP_CATEGORY_MARGIN = 0.20 # Margin above personal similarity for skip category classification # Safety & Operations MAX_DELETE_OPERATIONS_RATIO = 0.6 # Maximum delete operations ratio for safety @@ -318,9 +321,9 @@ class UnifiedCacheManager: class SkipDetector: - """Semantic-based content classifier using zero-shot classification with category descriptions.""" + """Binary content classifier: personal vs non-personal using semantic analysis.""" - TECHNICAL_CATEGORY_DESCRIPTIONS = [ + NON_PERSONAL_CATEGORY_DESCRIPTIONS = [ "programming language syntax, data types like string or integer, algorithm logic, function, method, programming class, object-oriented paradigm, variable scope, control flow, import, module, package, library, framework, recursion, iteration", "software design patterns, creational: singleton, factory, builder; structural: adapter, decorator, facade, proxy; behavioral: observer, strategy, command, mediator, chain of responsibility; abstract interface, polymorphism, composition", "error handling, exception, stack trace, TypeError, NullPointerException, IndexError, segmentation fault, core dump, stack overflow, runtime vs compile-time error, assertion failed, syntax error, null pointer dereference, memory leak, bug", @@ -341,9 +344,6 @@ class SkipDetector: "regex pattern, regular expression matching, groups, capturing, backslash escapes, metacharacters, wildcards, quantifiers, character classes, lookaheads, lookbehinds, alternation, anchors, word boundary, multiline flag, global search", "software testing, unit test, assertion, mock, stub, fixture, test suite, test case, verification, automated QA, validation framework, JUnit, pytest, Jest. Integration, end-to-end (E2E), functional, regression, acceptance testing", "cloud computing platforms, infrastructure as a service (IaaS), PaaS, AWS, Azure, GCP, compute instance, region, availability zone, elasticity, distributed system, virtual machine, container, serverless, Lambda, edge computing, CDN", - ] - - INSTRUCTION_CATEGORY_DESCRIPTIONS = [ "format the output as structured data. Return the answer as JSON with specific keys and values, or as YAML. Organize information into a CSV file or a database-style table with columns and rows. Present as a list of objects or an array.", "style the text presentation. Use markdown formatting like bullet points, a numbered list, or a task list. Organize content into a grid or tabular layout with proper alignment. Create a hierarchical structure with nested elements for clarity.", "adjust the response length. Make the answer shorter, more concise, brief, or condensed. Summarize the key points. Trim down the text to reduce the overall word count or meet a specific character limit. Be less verbose and more direct.", @@ -354,9 +354,6 @@ class SkipDetector: "continue the generated response. Keep going with the explanation or list. Provide more information and finish your thought. Complete the rest of the content or story. Proceed with the next steps. Do not stop until you have concluded.", "act as a specific persona or role. Respond as if you were a pirate, a scientist, or a travel guide. Adopt the character's voice, style, and knowledge base in your answer. Maintain the persona throughout the entire response.", "compare and contrast two or more topics. Explain the similarities and differences between A and B. Provide a detailed analysis of what they have in common and how they diverge. Create a table to highlight the key distinctions.", - ] - - PURE_MATH_CALCULATION_CATEGORY_DESCRIPTIONS = [ "perform a pure arithmetic calculation with explicit numbers. Solve, multiply, add, subtract, and divide. Compute a numeric expression following the order of operations (PEMDAS/BODMAS). What is 23 plus 456 minus 78 times 9 divided by 3?", "evaluate a mathematical expression containing numbers and operators, such as 2 plus 3 times 4 divided by 5. Solve this numerical problem and compute the final result. Simplify the arithmetic and show the final answer. Calculate 123 * 456.", "convert units between measurement systems with numeric values. Convert 100 kilometers to miles, 72 fahrenheit to celsius, or 5 feet 9 inches to centimeters. Change between metric and imperial for distance, weight, volume, or temperature.", @@ -367,9 +364,6 @@ class SkipDetector: "compute descriptive statistics for a dataset of numbers like 12, 15, 18, 20, 22. Calculate the mean, median, mode, average, and standard deviation. Find the variance, range, quartiles, and percentiles for a given sample distribution.", "calculate health and fitness metrics using a numeric formula. Compute the Body Mass Index (BMI) given a weight in pounds or kilograms and height in feet, inches, or meters. Find my basal metabolic rate (BMR) or target heart rate.", "calculate the time difference between two dates. How many days, hours, or minutes are between two points in time? Find the duration or elapsed time. Act as an age calculator for a birthday or find the time until a future anniversary.", - ] - - EXPLICIT_TRANSLATION_CATEGORY_DESCRIPTIONS = [ "translate the explicitly quoted text 'Hello, how are you?' to a foreign language like Spanish, French, or German. This is a translation instruction that includes the word 'translate' and the source text in quotes for direct conversion.", "how do you say a specific word or phrase in another language? For example, how do you say 'thank you', 'computer', or 'goodbye' in Japanese, Chinese, or Korean? This is a request for a direct translation of a common expression or term.", "convert a block of text or a paragraph from a source language to a target language. Translate the following content to Italian, Arabic, Portuguese, or Russian. This is a language conversion request for a larger piece of text provided.", @@ -380,9 +374,6 @@ class SkipDetector: "how do I say 'I am learning to code' in German? Convert this specific English phrase into its equivalent in another language. This is a request for a practical, conversational phrase translation for personal or professional use.", "translate this informal or slang expression to its colloquial equivalent in Spanish. How would you say 'What's up?' in Japanese in a casual context? This request focuses on capturing the correct tone and nuance of informal language.", "provide the formal and professional translation for 'Please find the attached document for your review' in French. Translate this business email phrase to German, ensuring the terminology and register are appropriate for a corporate context.", - ] - - GRAMMAR_PROOFREADING_CATEGORY_DESCRIPTIONS = [ "proofread the following text for errors. Here is my draft, please check it for typos and mistakes: 'Teh quick brown fox jumpped'. Review, revise, and correct any misspellings or grammatical issues you find in the provided passage.", "correct the grammar in this sentence: 'She don't like it'. Resolve grammatical issues like subject-verb agreement, incorrect verb tense, pronoun reference errors, or misplaced modifiers in the provided text. Address faulty sentence structure.", "check the spelling and punctuation in this passage. Please review the following text and correct any textual errors: 'its a beautiful day, isnt it'. Amend mistakes with commas, periods, apostrophes, quotation marks, colons, or capitalization.", @@ -395,7 +386,7 @@ class SkipDetector: "check my essay for conciseness and remove any redundancy. Help me edit this text to be more direct and to the point. Identify and eliminate wordiness, filler words, and repetitive phrases to strengthen the overall quality of the writing.", ] - CONVERSATIONAL_CATEGORY_DESCRIPTIONS = [ + PERSONAL_CATEGORY_DESCRIPTIONS = [ "discussing my family members, like my spouse, children, parents, or siblings. Mentioning relatives by name or role, such as my husband, wife, son, daughter, mother, or father. Sharing stories or asking questions about my family.", "expressing lasting personal feelings, core values, beliefs, or principles. My worldview, deeply held opinions, philosophy, or moral standards. Things I love, hate, or feel strongly about in life, such as my passion for animal welfare.", "describing my established personal hobbies, regular activities, or consistent interests. My passions and what I do in my leisure time, such as creative outlets like painting, sports like hiking, or other recreational pursuits I enjoy.", @@ -425,19 +416,11 @@ class SkipDetector: class SkipReason(Enum): SKIP_SIZE = "SKIP_SIZE" - SKIP_TECHNICAL = "SKIP_TECHNICAL" - SKIP_INSTRUCTION = "SKIP_INSTRUCTION" - SKIP_PURE_MATH = "SKIP_PURE_MATH" - SKIP_TRANSLATION = "SKIP_TRANSLATION" - SKIP_GRAMMAR_PROOFREAD = "SKIP_GRAMMAR_PROOFREAD" + SKIP_NON_PERSONAL = "SKIP_NON_PERSONAL" STATUS_MESSAGES = { SkipReason.SKIP_SIZE: "πŸ“ Message Length Out of Limits, skipping memory operations", - SkipReason.SKIP_TECHNICAL: "πŸ’» Technical Content Detected, skipping memory operations", - SkipReason.SKIP_INSTRUCTION: "πŸ’¬ Instruction Detected, skipping memory operations", - SkipReason.SKIP_PURE_MATH: "πŸ”’ Mathematical Calculation Detected, skipping memory operations", - SkipReason.SKIP_TRANSLATION: "🌐 Translation Request Detected, skipping memory operations", - SkipReason.SKIP_GRAMMAR_PROOFREAD: "πŸ“ Grammar/Proofreading Request Detected, skipping memory operations", + SkipReason.SKIP_NON_PERSONAL: "🚫 Non-Personal Content Detected, skipping memory operations", } def __init__(self, embedding_function: Callable[[Union[str, List[str]]], Union[np.ndarray, List[np.ndarray]]]): @@ -581,7 +564,7 @@ class SkipDetector: structured_lines = sum(1 for line in non_empty_lines if line.startswith((" ", "\t"))) if markup_in_lines / len(non_empty_lines) > 0.3: - return self.SkipReason.SKIP_TECHNICAL.value + return self.SkipReason.SKIP_NON_PERSONAL.value elif structured_lines / len(non_empty_lines) > 0.6: technical_keywords = ["function", "class", "import", "return", "const", "var", "let", "def"] if any(keyword in message.lower() for keyword in technical_keywords): @@ -613,7 +596,7 @@ class SkipDetector: """ Detect if a message should be skipped using two-stage detection: 1. Fast-path structural patterns (~95% confidence) - 2. Semantic classification (for remaining cases) + 2. Binary semantic classification (personal vs non-personal) Returns: Skip reason string if content should be skipped, None otherwise """ @@ -749,6 +732,34 @@ class LLMConsolidationService: def __init__(self, memory_system): self.memory_system = memory_system + async def _check_semantic_duplicate(self, content: str, existing_memories: List, user_id: str) -> Optional[str]: + """ + Check if content is semantically duplicate of existing memories using embeddings. + Returns the ID of duplicate memory if found, None otherwise. + """ + if not existing_memories: + return None + + try: + content_embedding = await self.memory_system._generate_embeddings(content, user_id) + + for memory in existing_memories: + if not memory.content or len(memory.content.strip()) < Constants.MIN_MESSAGE_CHARS: + continue + + memory_embedding = await self.memory_system._generate_embeddings(memory.content, user_id) + + similarity = float(np.dot(content_embedding, memory_embedding)) + + if similarity >= Constants.DEDUPLICATION_SIMILARITY_THRESHOLD: + logger.info(f"πŸ” Semantic duplicate detected: similarity={similarity:.3f} with memory {memory.id}") + return str(memory.id) + + return None + except Exception as e: + logger.warning(f"⚠️ Semantic duplicate check failed: {str(e)}") + return None + def _filter_consolidation_candidates(self, similarities: List[Dict[str, Any]]) -> Tuple[List[Dict[str, Any]], str]: """Filter consolidation candidates by threshold and return candidates with threshold info.""" consolidation_threshold = self.memory_system._get_retrieval_threshold(is_consolidation=True) @@ -841,7 +852,31 @@ class LLMConsolidationService: ) return [] - valid_operations = [op.model_dump() for op in operations if op.validate_operation(existing_memory_ids)] + deduplicated_operations = [] + seen_contents = set() + seen_update_ids = set() + + for op in operations: + if not op.validate_operation(existing_memory_ids): + continue + + if op.operation == Models.MemoryOperationType.UPDATE and op.id in seen_update_ids: + logger.info(f"⏭️ Skipping duplicate UPDATE for memory {op.id} in LLM response") + continue + + if op.operation in [Models.MemoryOperationType.CREATE, Models.MemoryOperationType.UPDATE]: + normalized_content = op.content.strip().lower() + if normalized_content in seen_contents: + op_type = "CREATE" if op.operation == Models.MemoryOperationType.CREATE else f"UPDATE {op.id}" + logger.info(f"⏭️ Skipping duplicate {op_type} in LLM response: {self.memory_system._truncate_content(op.content)}") + continue + seen_contents.add(normalized_content) + + if op.operation == Models.MemoryOperationType.UPDATE: + seen_update_ids.add(op.id) + deduplicated_operations.append(op.model_dump()) + + valid_operations = deduplicated_operations if valid_operations: create_count = sum(1 for op in valid_operations if op.get("operation") == Models.MemoryOperationType.CREATE.value) @@ -856,6 +891,41 @@ class LLMConsolidationService: return valid_operations + async def _deduplicate_operations( + self, operations: List, current_memories: List, user_id: str, operation_type: str, delete_operations: Optional[List] = None + ) -> List: + """ + Deduplicate operations against existing memories using semantic similarity. + For UPDATE operations, preserves enriched content and deletes the duplicate. + """ + deduplicated = [] + + for operation in operations: + memories_to_check = current_memories + if operation_type == "UPDATE": + memories_to_check = [m for m in current_memories if str(m.id) != operation.id] + + duplicate_id = await self._check_semantic_duplicate(operation.content, memories_to_check, user_id) + + if duplicate_id: + if operation_type == "UPDATE" and delete_operations is not None: + logger.info( + f"πŸ”„ UPDATE creates duplicate: keeping enriched content from memory {operation.id}, " f"deleting duplicate memory {duplicate_id}" + ) + deduplicated.append(operation) + delete_operations.append(Models.MemoryOperation(operation=Models.MemoryOperationType.DELETE, content="", id=duplicate_id)) + else: + logger.info( + f"⏭️ Skipping duplicate {operation_type}: " + f"{self.memory_system._truncate_content(operation.content)} " + f"(matches memory {duplicate_id})" + ) + continue + + deduplicated.append(operation) + + return deduplicated + async def execute_memory_operations(self, operations: List[Dict[str, Any]], user_id: str, emitter: Optional[Callable] = None) -> Tuple[int, int, int, int]: """Execute consolidation operations with simplified tracking.""" if not operations or not user_id: @@ -898,6 +968,23 @@ class LLMConsolidationService: except Exception as e: logger.warning(f"⚠️ Failed to fetch memories for DELETE preview: {str(e)}") + if operations_by_type["CREATE"] or operations_by_type["UPDATE"]: + try: + current_memories = await self.memory_system._get_user_memories(user_id) + + if operations_by_type["CREATE"]: + operations_by_type["CREATE"] = await self._deduplicate_operations( + operations_by_type["CREATE"], current_memories, user_id, operation_type="CREATE" + ) + + if operations_by_type["UPDATE"]: + operations_by_type["UPDATE"] = await self._deduplicate_operations( + operations_by_type["UPDATE"], current_memories, user_id, operation_type="UPDATE", delete_operations=operations_by_type["DELETE"] + ) + + except Exception as e: + logger.warning(f"⚠️ Semantic deduplication check failed, proceeding with original operations: {str(e)}") + for operation_type, ops in operations_by_type.items(): if not ops: continue @@ -1031,6 +1118,7 @@ class Filter: self._shutdown_event = asyncio.Event() self._embedding_function = None + self._embedding_dimension = None self._skip_detector = None self._llm_reranking_service = LLMRerankingService(self) @@ -1133,12 +1221,32 @@ class Filter: """Compute SHA256 hash for text caching.""" return hashlib.sha256(text.encode()).hexdigest() + def _detect_embedding_dimension(self) -> None: + """Detect embedding dimension by generating a test embedding.""" + try: + test_embedding = self._embedding_function(["dummy"], prefix=None, user=None) + if isinstance(test_embedding, list): + test_embedding = test_embedding[0] + self._embedding_dimension = np.squeeze(test_embedding).shape[0] + logger.info(f"🎯 Detected embedding dimension: {self._embedding_dimension}") + except Exception as e: + raise RuntimeError(f"Failed to detect embedding dimension: {str(e)}") + def _normalize_embedding(self, embedding: Union[List[float], np.ndarray]) -> np.ndarray: - """Normalize embedding vector.""" + """Normalize embedding vector and ensure 1D shape.""" if isinstance(embedding, list): embedding = np.array(embedding, dtype=np.float16) else: embedding = embedding.astype(np.float16) + + embedding = np.squeeze(embedding) + + if embedding.ndim != 1: + raise ValueError(f"Embedding must be 1D after squeeze, got shape {embedding.shape}") + + if self._embedding_dimension and embedding.shape[0] != self._embedding_dimension: + raise ValueError(f"Embedding must have {self._embedding_dimension} dimensions, got {embedding.shape[0]}") + norm = np.linalg.norm(embedding) return embedding / norm if norm > 0 else embedding @@ -1185,7 +1293,7 @@ class Filter: raw_embeddings = await loop.run_in_executor(None, self._embedding_function, uncached_texts, None, user) if isinstance(raw_embeddings, list) and len(raw_embeddings) > 0: - if isinstance(raw_embeddings[0], list): + if isinstance(raw_embeddings[0], (list, np.ndarray)): new_embeddings = [self._normalize_embedding(emb) for emb in raw_embeddings] else: new_embeddings = [self._normalize_embedding(raw_embeddings)] @@ -1199,7 +1307,8 @@ class Filter: result_embeddings[original_idx] = embedding if is_single: - logger.info("πŸ“₯ User message embedding: cache hit" if not uncached_texts else "πŸ’Ύ User message embedding: generated and cached") + if uncached_texts: + logger.info("πŸ’Ύ User message embedding: generated and cached") return result_embeddings[0] else: valid_count = sum(1 for emb in result_embeddings if emb is not None) diff --git a/requirements.txt b/requirements.txt index 910811c..6b04c53 100644 --- a/requirements.txt +++ b/requirements.txt @@ -4,4 +4,4 @@ pydantic numpy tiktoken black -isort \ No newline at end of file +isort From b5a487209670ecf94223520303d8435c7ba376c2 Mon Sep 17 00:00:00 2001 From: mtayfur Date: Mon, 27 Oct 2025 00:57:26 +0300 Subject: [PATCH 4/4] =?UTF-8?q?=F0=9F=93=9D=20(memory=5Fsystem):=20clarify?= =?UTF-8?q?=20and=20strengthen=20intent=20filtering=20and=20memory=20conso?= =?UTF-8?q?lidation=20guidelines?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Expand and clarify the "Filter for Intent" rule to ensure only direct, personally significant facts are stored, explicitly excluding messages where the user's primary intent is instructional, technical, or analytical. Update processing and decision frameworks to reinforce selectivity based on user intent. Revise and annotate examples to demonstrate correct application of the new rules, making it clear that requests for advice, recommendations, or technical tasks are ignored. These changes improve the precision and reliability of memory consolidation, reducing the risk of storing irrelevant or transient information. --- memory_system.py | 28 +++++++++++++++------------- 1 file changed, 15 insertions(+), 13 deletions(-) diff --git a/memory_system.py b/memory_system.py index 394a634..ebf1277 100644 --- a/memory_system.py +++ b/memory_system.py @@ -73,7 +73,7 @@ class Prompts: MEMORY_CONSOLIDATION = f"""You are the Memory System Consolidator, a specialist in creating precise user memories. ## OBJECTIVE -Build precise memories of the user's personal narrative with factual, temporal statements. +Your goal is to build precise memories of the user's personal narrative with factual, temporal statements. ## AVAILABLE OPERATIONS - CREATE: For new, personal facts. Must be semantically and temporally enhanced. @@ -83,6 +83,11 @@ Build precise memories of the user's personal narrative with factual, temporal s ## PROCESSING GUIDELINES - Personal Facts Only: Store only significant facts with lasting relevance to the user's life and identity. Exclude transient situations, questions, general knowledge, casual mentions, or momentary states. +- **Filter for Intent:** You MUST SKIP if the user's primary intent is instructional, technical, or analytical, even if the message contains personal details. This includes requests to: + - Rewrite, revise, translate, or proofread a block of text (e.g., "revise this review for me"). + - Answer a general knowledge, math, or technical question. + - Explain a concept, perform a calculation, or act as a persona. + **Only store facts when the user is *directly stating* them as part of a personal narrative, not when providing them as content for a task.** - Maintain Temporal Accuracy: - Capture Dates: Record temporal information when explicitly stated or clearly derivable. Convert relative references (last month, yesterday) to specific dates. - Preserve History: Transform superseded facts into past-tense statements with defined time boundaries. @@ -93,14 +98,13 @@ Build precise memories of the user's personal narrative with factual, temporal s - Retroactive Enrichment: If a name is provided for prior entity, UPDATE only if substantially valuable. - Ensure Memory Quality: - High Bar for Creation: Only CREATE memories for significant life facts, relationships, events, or core personal attributes. Skip trivial details or passing interests. - - Contextual Completeness: Combine related information into cohesive statements. Group connected facts (same topic, person, event, or timeframe) into single memories rather than fragmenting. Include supporting details while respecting boundaries. Only combine directly related facts. Avoid bare statements and never merge unrelated information. - Mandatory Semantic Enhancement: Enhance entities with descriptive categorical nouns for better retrieval. - Verify Nouns/Pronouns: Link pronouns (he, she, they) and nouns to specific entities. - First-Person Format: Write all memories in English from the user's perspective. ## DECISION FRAMEWORK -- Selectivity: Verify the user is stating a direct, personally significant fact with lasting importance. If not, SKIP. Never create duplicate memories. Skip momentary events or casual mentions. Be conservative with CREATE and UPDATE operations. -- Strategy: Strongly prioritize enriching existing memories over creating new ones. Analyze the message holistically to identify naturally connected facts that should be captured together. When facts share connections (same person, event, situation, or causal relationship), combine them into a unified memory that preserves the complete picture. Each memory should be self-contained and meaningful. +- Selectivity: Verify the user's *primary intent* is to state a direct, personally significant fact with lasting importance. If the intent is instructional, analytical, or a general question, SKIP. Never create duplicate memories. Skip momentary events or casual mentions. Be conservative with CREATE and UPDATE operations. +- Strategy: Strongly prioritize enriching existing memories over creating new ones. Analyze the message holistically to identify naturally connected facts (same person, event, or timeframe) and combine them into a unified, cohesive memory rather than fragmenting them. Each memory must be self-contained and **never** merge unrelated information. - Execution: For new significant facts, use CREATE. For simple attribute changes, use UPDATE only if it meaningfully improves the memory. For significant changes, use UPDATE to make the old memory historical, then CREATE the new one. For contradictions, use DELETE. ## EXAMPLES (Assumes Current Date: September 15, 2025) @@ -115,37 +119,37 @@ Explanation: Multiple facts about the same person (Sarah's active lifestyle, lov Message: "My daughter Emma just turned 12. We adopted a dog named Max for her 11th birthday. What should I give her for her 12th birthday?" Memories: [id:mem-002] My daughter Emma is 10 years old [noted at March 20 2024] [id:mem-101] I have a golden retriever [noted at September 20 2024] Return: {{"ops": [{{"operation": "UPDATE", "id": "mem-002", "content": "My daughter Emma turned 12 years old in September 2025"}}, {{"operation": "UPDATE", "id": "mem-101", "content": "I have a golden retriever named Max that was adopted in September 2024 as a birthday gift for my daughter Emma when she turned 11"}}]}} -Explanation: Dog memory enriched with related context (Emma, birthday gift, age 11) and temporal anchoring (September 2024) - all semantically connected to the same event and relationship. +Explanation: Dog memory enriched with related context (Emma, birthday gift, age 11) and temporal anchoring (September 2024). The instructional question ("What should I give her...?") is ignored as per the 'Filter for Intent' rule. ### Example 3 Message: "Can you recommend some good tapas restaurants in Barcelona? I moved here from Madrid last month." Memories: [id:mem-005] I live in Madrid Spain [noted at June 12 2025] Return: {{"ops": [{{"operation": "UPDATE", "id": "mem-005", "content": "I lived in Madrid Spain until August 2025"}}, {{"operation": "CREATE", "id": "", "content": "I moved to Barcelona Spain in August 2025"}}]}} -Explanation: Relocation is a significant life event with lasting impact. "Exploring the city" and "adjusting" are transient states and excluded. +Explanation: Relocation is a significant life event. The request for recommendations is instructional and is ignored. ### Example 4 Message: "My wife Sofia and I just got married in August. What are some good honeymoon destinations?" Memories: [id:mem-008] I am single [noted at January 5 2025] Return: {{"ops": [{{"operation": "DELETE", "id": "mem-008", "content": ""}}, {{"operation": "CREATE", "id": "", "content": "I married Sofia in August 2025 and she is now my wife"}}]}} -Explanation: Marriage is an enduring life event. Wife's name and marriage date are lasting facts combined naturally. "Planning honeymoon" is a transient activity and excluded. +Explanation: Marriage is an enduring life event. The instructional question ("What are some good honeymoon destinations?") is ignored. ### Example 5 Message: "Β‘Hola! Me mudΓ© de Madrid a Barcelona el mes pasado y me casΓ© con mi novia SofΓ­a en agosto. ΒΏMe puedes recomendar un buen restaurante para celebrar?" Memories: [id:mem-005] I live in Madrid Spain [noted at June 12 2025] [id:mem-006] I am dating Sofia [noted at February 10 2025] [id:mem-008] I am single [noted at January 5 2025] Return: {{"ops": [{{"operation": "UPDATE", "id": "mem-005", "content": "I lived in Madrid Spain until August 2025"}}, {{"operation": "DELETE", "id": "mem-008", "content": ""}}, {{"operation": "UPDATE", "id": "mem-006", "content": "I moved to Barcelona Spain and married my girlfriend Sofia in August 2025, who is now my wife"}}]}} -Explanation: The user's move and marriage are significant, related life events that occurred in the same month. They are consolidated into a single, cohesive memory that enriches the existing relationship context. +Explanation: The user's move and marriage are significant, related life events. They are consolidated into a single memory. The request for a recommendation is ignored. ### Example 6 Message: "I'm feeling stressed about work this week and looking for some relaxation tips. I have a big presentation coming up on Friday." Memories: [] Return: {{"ops": []}} -Explanation: Temporary stress, seeking tips, and upcoming presentation are all transient situations without lasting personal significance. Nothing to store. +Explanation: Transient state (stress) and a request for information (relaxation tips). The primary intent is instructional/analytical, and the facts (presentation) are not significant, lasting personal narrative. Nothing to store. """ MEMORY_RERANKING = f"""You are the Memory Relevance Analyzer. ## OBJECTIVE -Select relevant memories to personalize the response, prioritizing direct connections and supporting context. +Your goal is to analyze the user's message and select the most relevant memories to personalize the AI's response. Prioritize direct connections and supporting context. ## RELEVANCE CATEGORIES - Direct: Memories explicitly about the query topic, people, or domain. @@ -154,9 +158,8 @@ Select relevant memories to personalize the response, prioritizing direct connec ## SELECTION FRAMEWORK - Prioritize Current Info: Give current facts higher relevance than historical ones unless the query is about the past or historical context directly informs the current situation. -- Hierarchy: Prioritize Direct β†’ Contextual β†’ Background. +- Hierarchy: Prioritize topic matches first (Direct), then context that enhances the response (Contextual), and finally general background (Background). - Ordering: Order IDs by relevance, most relevant first. -- Standard: Prioritize topic matches, then context that enhances the response. - Maximum Limit: Return up to {Constants.MAX_MEMORIES_PER_RETRIEVAL} memory IDs. ## EXAMPLES (Assumes Current Date: September 15, 2025) @@ -186,7 +189,6 @@ Return: {{"ids": []}} Explanation: Query seeks general technical explanation without personal context. Job and family information don't affect how quantum computing concepts should be explained. """ - class Models: """Container for all Pydantic models used in the memory system."""