File size: 18,927 Bytes
a54fd97 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 | """
Memory Builder
Stage 1: Semantic Structured Compression (Section 3.1)
& Stage 2: Online Semantic Synthesis (Section 3.2)
Implements:
- Implicit semantic density gating: Φ_gate(W) → {m_k} (filters low-density windows)
- Sliding window processing for dialogue segmentation
- Generates compact memory units with resolved coreferences and absolute timestamps
"""
from typing import List, Optional
from models.memory_entry import MemoryEntry, Dialogue
from utils.llm_client import LLMClient
from database.vector_store import VectorStore
import config
import json
import asyncio
import concurrent.futures
from functools import partial
class MemoryBuilder:
"""
Memory Builder - Semantic Structured Compression (Section 3.1)
Core Functions:
1. Sliding window segmentation
2. Implicit semantic density gating: Φ_gate(W) → {m_k}
3. Multi-view indexing: I(m_k) = {s_k, l_k, r_k}
4. Intra-session consolidation during write (Section 3.2): by generating enough memory entries to ensure ALL information is captured
"""
def __init__(
self,
llm_client: LLMClient,
vector_store: VectorStore,
window_size: int = None,
overlap_size: Optional[int] = None,
enable_parallel_processing: bool = True,
max_parallel_workers: int = 3,
fail_on_extraction_error: Optional[bool] = None,
):
self.llm_client = llm_client
self.vector_store = vector_store
self.window_size = window_size or config.WINDOW_SIZE
self.overlap_size = overlap_size if overlap_size is not None else getattr(config, 'OVERLAP_SIZE', 0)
# step_size is how far the window advances each iteration; overlap retains
# the last overlap_size dialogues so the next window has continuity context
self.step_size = max(1, self.window_size - self.overlap_size)
# Use config values as default if not explicitly provided
self.enable_parallel_processing = enable_parallel_processing if enable_parallel_processing is not None else getattr(config, 'ENABLE_PARALLEL_PROCESSING', True)
self.max_parallel_workers = max_parallel_workers if max_parallel_workers is not None else getattr(config, 'MAX_PARALLEL_WORKERS', 4)
# Dialogue buffer
self.dialogue_buffer: List[Dialogue] = []
self.processed_count = 0
self.fail_on_extraction_error = (
fail_on_extraction_error
if fail_on_extraction_error is not None
else getattr(config, "FAIL_ON_EXTRACTION_ERROR", False)
)
# Previous window entries (for context)
self.previous_entries: List[MemoryEntry] = []
def add_dialogue(self, dialogue: Dialogue, auto_process: bool = True):
"""
Add a dialogue to the buffer
"""
self.dialogue_buffer.append(dialogue)
# Auto process
if auto_process and len(self.dialogue_buffer) >= self.window_size:
self.process_window()
def add_dialogues(self, dialogues: List[Dialogue], auto_process: bool = True):
"""
Batch add dialogues with optional parallel processing
"""
if self.enable_parallel_processing and len(dialogues) > self.window_size * 2:
# Use parallel processing for large batches
self.add_dialogues_parallel(dialogues)
else:
# Use sequential processing for smaller batches
for dialogue in dialogues:
self.add_dialogue(dialogue, auto_process=False)
# Process complete windows
if auto_process:
while len(self.dialogue_buffer) >= self.window_size:
self.process_window()
def add_dialogues_parallel(self, dialogues: List[Dialogue]):
"""
Add dialogues using parallel processing for better performance
"""
# Snapshot pre-existing buffer items so the fallback can restore them
# if the buffer is cleared mid-way through parallel processing
pre_existing = list(self.dialogue_buffer)
windows_to_process = []
try:
# Add all dialogues to buffer first
self.dialogue_buffer.extend(dialogues)
# Group into windows using step_size so that each window retains
# overlap_size dialogues of context from the previous window
pos = 0
while pos + self.window_size <= len(self.dialogue_buffer):
window = self.dialogue_buffer[pos:pos + self.window_size]
windows_to_process.append(window)
pos += self.step_size
# Add remaining dialogues as a smaller batch (no need to process separately)
remaining = self.dialogue_buffer[pos:]
if remaining:
windows_to_process.append(remaining)
self.dialogue_buffer = [] # Clear buffer since we're processing all
if windows_to_process:
print(f"\n[Parallel Processing] Processing {len(windows_to_process)} batches in parallel with {self.max_parallel_workers} workers")
print(f"Batch sizes: {[len(w) for w in windows_to_process]}")
# Process all windows/batches in parallel (including remaining dialogues)
self._process_windows_parallel(windows_to_process)
except Exception as e:
print(f"[Parallel Processing] Failed: {e}. Falling back to sequential processing...")
# Fallback: overlapping windows cannot be re-stacked naively.
# If the buffer was cleared (exception after line 107), restore the full
# original state: pre-existing items that were already in the buffer
# PLUS the new dialogues we were asked to process.
# If the buffer was NOT cleared (exception before line 107), it already
# contains pre_existing + dialogues, so leave it as-is.
if not self.dialogue_buffer:
self.dialogue_buffer = pre_existing + list(dialogues)
# process_window() uses step_size, so overlap is handled correctly here
while len(self.dialogue_buffer) >= self.window_size:
self.process_window()
def process_window(self):
"""
Process current window dialogues - Core logic
"""
if not self.dialogue_buffer:
return
# Extract window; advance by step_size to retain overlap_size dialogues
# at the tail so the next window has continuity context
window = self.dialogue_buffer[:self.window_size]
self.dialogue_buffer = self.dialogue_buffer[self.step_size:]
print(f"\nProcessing window: {len(window)} dialogues (processed {self.processed_count} so far)")
# Call LLM to generate memory entries
entries = self._generate_memory_entries(window)
self.processed_count += len(window)
# Store to database
if entries:
self.vector_store.add_entries(entries)
self.previous_entries = entries # Save as context
print(f"Generated {len(entries)} memory entries")
def process_remaining(self):
"""
Process remaining dialogues (fallback method, normally handled in parallel)
"""
if self.dialogue_buffer:
print(f"\nProcessing remaining dialogues: {len(self.dialogue_buffer)} (fallback mode)")
entries = self._generate_memory_entries(self.dialogue_buffer)
self.processed_count += len(self.dialogue_buffer)
if entries:
self.vector_store.add_entries(entries)
self.dialogue_buffer = []
print(f"Generated {len(entries)} memory entries")
def _generate_memory_entries(self, dialogues: List[Dialogue]) -> List[MemoryEntry]:
"""
Implicit Semantic Density Gating (Section 3.1)
Φ_gate(W) → {m_k}, generates compact memory units from dialogue window
"""
# Build dialogue text
dialogue_text = "\n".join([str(d) for d in dialogues])
dialogue_ids = [d.dialogue_id for d in dialogues]
# Build context
context = ""
if self.previous_entries:
context = "\n[Previous Window Memory Entries (for reference to avoid duplication)]\n"
for entry in self.previous_entries[:3]: # Only show first 3
context += f"- {entry.lossless_restatement}\n"
# Build prompt
prompt = self._build_extraction_prompt(dialogue_text, dialogue_ids, context)
# Call LLM
messages = [
{
"role": "system",
"content": "You are a professional information extraction assistant, skilled at extracting structured, unambiguous information from conversations. You must output valid JSON format."
},
{
"role": "user",
"content": prompt
}
]
# Retry up to 3 times if parsing fails
max_retries = 3
last_exception = None
for attempt in range(max_retries):
try:
# Use JSON format if configured
response_format = None
if hasattr(config, 'USE_JSON_FORMAT') and config.USE_JSON_FORMAT:
response_format = {"type": "json_object"}
response = self.llm_client.chat_completion(
messages,
temperature=0.1,
response_format=response_format,
max_tokens=2600,
)
# Parse response
entries = self._parse_llm_response(response, dialogue_ids)
return entries
except Exception as e:
last_exception = e
if attempt < max_retries - 1:
print(f"Attempt {attempt + 1}/{max_retries} failed to parse LLM response: {e}")
print(f"Retrying...")
else:
print(f"All {max_retries} attempts failed to parse LLM response: {e}")
print(f"Raw response: {response[:500] if 'response' in locals() else 'No response'}")
if self.fail_on_extraction_error:
raise RuntimeError(
f"Failed to generate memory entries for dialogue ids {dialogue_ids}"
) from last_exception
return []
def _build_extraction_prompt(
self,
dialogue_text: str,
dialogue_ids: List[int],
context: str
) -> str:
"""
Build LLM extraction prompt
"""
return f"""
Your task is to extract all valuable information from the following dialogues and convert them into structured memory entries.
{context}
[Current Window Dialogues]
{dialogue_text}
[Requirements]
1. **Complete Coverage With Concise Synthesis**: Preserve important information from the dialogues while merging tightly related facts into semantically dense entries instead of many small entries
2. **Force Disambiguation**: Absolutely PROHIBIT using pronouns (he, she, it, they, this, that) and relative time (yesterday, today, last week, tomorrow)
3. **Lossless Information**: Each entry's lossless_restatement must be a complete, independent, understandable sentence
4. **Precise Extraction**:
- keywords: Core keywords (names, places, entities, topic words)
- timestamp: Absolute time in ISO 8601 format (if explicit time mentioned in dialogue)
- location: Specific location name (if mentioned)
- persons: All person names mentioned
- entities: Companies, products, organizations, etc.
- topic: The topic of this information
[Output Format]
Return a JSON object with an "entries" array, where each element is a memory entry:
```json
{{
"entries": [
{{
"lossless_restatement": "Complete unambiguous restatement (must include all subjects, objects, time, location, etc.)",
"keywords": ["keyword1", "keyword2", ...],
"timestamp": "YYYY-MM-DDTHH:MM:SS or null",
"location": "location name or null",
"persons": ["name1", "name2", ...],
"entities": ["entity1", "entity2", ...],
"topic": "topic phrase"
}},
...
]
}}
```
[Example]
Dialogues:
[2025-11-15T14:30:00] Alice: Bob, let's meet at Starbucks tomorrow at 2pm to discuss the new product
[2025-11-15T14:31:00] Bob: Okay, I'll prepare the materials
Output:
```json
{{
"entries": [
{{
"lossless_restatement": "Alice suggested at 2025-11-15T14:30:00 to meet with Bob at Starbucks on 2025-11-16T14:00:00 to discuss the new product.",
"keywords": ["Alice", "Bob", "Starbucks", "new product", "meeting"],
"timestamp": "2025-11-16T14:00:00",
"location": "Starbucks",
"persons": ["Alice", "Bob"],
"entities": ["new product"],
"topic": "Product discussion meeting arrangement"
}},
{{
"lossless_restatement": "Bob agreed to attend the meeting and committed to prepare relevant materials.",
"keywords": ["Bob", "prepare materials", "agree"],
"timestamp": null,
"location": null,
"persons": ["Bob"],
"entities": [],
"topic": "Meeting preparation confirmation"
}}
]
}}
```
Now process the above dialogues. Return ONLY the JSON object, no other explanations.
"""
def _parse_llm_response(
self,
response: str,
dialogue_ids: List[int]
) -> List[MemoryEntry]:
"""
Parse LLM response to MemoryEntry list
"""
# Extract JSON
data = self.llm_client.extract_json(response)
if isinstance(data, dict):
data = data.get("entries")
if not isinstance(data, list):
raise ValueError(f"Expected JSON array or object with 'entries' but got: {type(data)}")
entries = []
for item in data:
# Create MemoryEntry
entry = MemoryEntry(
lossless_restatement=item["lossless_restatement"],
keywords=item.get("keywords", []),
timestamp=item.get("timestamp"),
location=item.get("location"),
persons=item.get("persons", []),
entities=item.get("entities", []),
topic=item.get("topic")
)
entries.append(entry)
return entries
def _process_windows_parallel(self, windows: List[List[Dialogue]]):
"""
Process multiple windows in parallel using ThreadPoolExecutor
"""
all_entries = []
# Use ThreadPoolExecutor for parallel processing
with concurrent.futures.ThreadPoolExecutor(max_workers=self.max_parallel_workers) as executor:
# Submit all window processing tasks
future_to_window = {}
for i, window in enumerate(windows):
dialogue_ids = [d.dialogue_id for d in window]
future = executor.submit(self._generate_memory_entries_worker, window, dialogue_ids, i+1)
future_to_window[future] = (window, i+1)
# Collect results as they complete
for future in concurrent.futures.as_completed(future_to_window):
window, window_num = future_to_window[future]
try:
entries = future.result()
all_entries.extend(entries)
print(f"[Parallel Processing] Window {window_num} completed: {len(entries)} entries")
except Exception as e:
print(f"[Parallel Processing] Window {window_num} failed: {e}")
# Store all entries to database in batch
if all_entries:
print(f"\n[Parallel Processing] Storing {len(all_entries)} entries to database...")
self.vector_store.add_entries(all_entries)
self.processed_count += sum(len(window) for window in windows)
# Update previous entries (use last window's entries for context)
if all_entries:
self.previous_entries = all_entries[-10:] # Keep last 10 entries for context
print(f"[Parallel Processing] Completed processing {len(windows)} windows")
def _generate_memory_entries_worker(self, window: List[Dialogue], dialogue_ids: List[int], window_num: int) -> List[MemoryEntry]:
"""
Worker function for parallel processing of a single batch (full window or remaining dialogues)
"""
batch_size = len(window)
batch_type = "full window" if batch_size == self.window_size else f"remaining batch"
print(f"[Worker {window_num}] Processing {batch_type} with {batch_size} dialogues")
# Build dialogue text
dialogue_text = "\n".join([str(d) for d in window])
# Build context (shared across all workers - this is fine for parallel processing)
context = ""
if self.previous_entries:
context = "\n[Previous Window Memory Entries (for reference to avoid duplication)]\n"
for entry in self.previous_entries[:3]: # Only show first 3
context += f"- {entry.lossless_restatement}\n"
# Build prompt
prompt = self._build_extraction_prompt(dialogue_text, dialogue_ids, context)
# Call LLM
messages = [
{
"role": "system",
"content": "You are a professional information extraction assistant, skilled at extracting structured, unambiguous information from conversations. You must output valid JSON format."
},
{
"role": "user",
"content": prompt
}
]
# Retry up to 3 times if parsing fails
max_retries = 3
for attempt in range(max_retries):
try:
# Use JSON format if configured
response_format = None
if hasattr(config, 'USE_JSON_FORMAT') and config.USE_JSON_FORMAT:
response_format = {"type": "json_object"}
response = self.llm_client.chat_completion(
messages,
temperature=0.1,
response_format=response_format,
max_tokens=2600,
)
# Parse response
entries = self._parse_llm_response(response, dialogue_ids)
print(f"[Worker {window_num}] Generated {len(entries)} entries")
return entries
except Exception as e:
if attempt < max_retries - 1:
print(f"[Worker {window_num}] Attempt {attempt + 1}/{max_retries} failed: {e}. Retrying...")
else:
print(f"[Worker {window_num}] All {max_retries} attempts failed: {e}")
return []
|