File size: 42,369 Bytes
5e028bf | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 | import uuid
import re
import copy
import concurrent
import logging
import json
import threading
from datetime import datetime, timedelta
from typing import Any, Dict, Literal, Optional, List, Tuple, Union
from pydantic import ValidationError
from lightmem.configs.base import BaseMemoryConfigs
from lightmem.factory.pre_compressor.factory import PreCompressorFactory
from lightmem.factory.topic_segmenter.factory import TopicSegmenterFactory
from lightmem.factory.memory_manager.factory import MemoryManagerFactory
from lightmem.factory.text_embedder.factory import TextEmbedderFactory
from lightmem.factory.retriever.contextretriever.factory import ContextRetrieverFactory
from lightmem.factory.retriever.embeddingretriever.factory import EmbeddingRetrieverFactory
from lightmem.factory.retriever.embeddingretriever.qdrant import QdrantConfig
from lightmem.factory.memory_buffer.sensory_memory import SenMemBufferManager
from lightmem.factory.memory_buffer.short_term_memory import ShortMemBufferManager
from lightmem.memory.utils import MemoryEntry, assign_sequence_numbers_with_timestamps, save_memory_entries,convert_extraction_results_to_memory_entries,normalize_extraction_prompts,process_extraction_results
from lightmem.memory.prompts import METADATA_GENERATE_PROMPT, UPDATE_PROMPT
from lightmem.configs.logging.utils import get_logger
GLOBAL_TOPIC_IDX = 0
GLOBAL_LAST_SUMMARY_TIME = None
class MessageNormalizer:
_SESSION_RE = re.compile(
r'(?P<date>\d{4}[/-]\d{1,2}[/-]\d{1,2})\s*\((?P<weekday>[^)]+)\)\s*(?P<time>\d{1,2}:\d{2}(?::\d{2})?)'
)
def __init__(self, offset_ms: int = 1000):
self.last_timestamp_map: Dict[str, datetime] = {}
self.offset = timedelta(milliseconds=offset_ms)
def _parse_session_timestamp(self, raw_ts: str) -> Tuple[datetime, str]:
"""
Parse a session-level timestamp and return (base_datetime, weekday).
Supports formats like "2023/05/20 (Sat) 00:44" (also accepts '-' as separator, and optional seconds).
Raises ValueError if parsing fails.
"""
m = self._SESSION_RE.search(raw_ts)
if m:
date_str = m.group('date').replace('-', '/')
time_str = m.group('time')
weekday = m.group('weekday')
fmt = "%Y/%m/%d %H:%M:%S" if time_str.count(':') == 2 else "%Y/%m/%d %H:%M"
base_dt = datetime.strptime(f"{date_str} {time_str}", fmt)
return base_dt, weekday
try:
dt = datetime.fromisoformat(raw_ts)
return dt, dt.strftime("%a")
except Exception as e:
raise ValueError(f"{str(e)}: Failed to parse session time format: '{raw_ts}'. Expected something like '2023/05/20 (Sat) 00:44'")
def normalize_messages(self, messages: Any) -> List[Dict[str, Any]]:
"""
Accepts str / dict / list[dict]:
- If str -> treated as a single user message (if 'time_stamp' is required, use dict form)
- If dict -> single message
- If list -> multiple messages (each must be a dict and contain 'time_stamp')
Returns: List[Dict] (each item is a copied and enriched message)
"""
# Normalize input into a list
if isinstance(messages, dict):
messages_list = [messages]
elif isinstance(messages, list):
messages_list = messages
elif isinstance(messages, str):
raise ValueError("Please provide messages as dict or list[dict], and ensure each dict contains a 'time_stamp' field (session-level).")
else:
raise ValueError("messages must be dict or list[dict] (or str, but not recommended).")
enriched_list: List[Dict[str, Any]] = []
for msg in messages_list:
if not isinstance(msg, dict):
raise ValueError("Each item in messages list must be a dict.")
raw_ts = msg.get("time_stamp")
if not raw_ts:
raise ValueError("Each message should contain a 'time_stamp' field (e.g., '2023/05/20 (Sat) 00:44').")
base_dt, weekday = self._parse_session_timestamp(raw_ts)
# Maintain incrementing time based on raw_ts as session key
last_dt = self.last_timestamp_map.get(raw_ts)
if last_dt is None:
new_dt = base_dt
else:
new_dt = last_dt + self.offset
self.last_timestamp_map[raw_ts] = new_dt
enriched = copy.deepcopy(msg)
enriched["session_time"] = raw_ts
enriched["time_stamp"] = new_dt.isoformat(timespec="milliseconds")
enriched["weekday"] = weekday
enriched_list.append(enriched)
return enriched_list
class LightMemory:
def __init__(self, config: BaseMemoryConfigs = BaseMemoryConfigs()):
"""
Initialize a LightMemory instance.
This constructor initializes various memory-related components based on the provided configuration (`config`),
including the memory manager, optional pre-compressor, optional topic segmenter, text embedder,
and retrievers based on the configured strategies.
This design supports flexible extension of the memory system, making it easy to integrate
different processing and retrieval capabilities.
Args:
config (BaseMemoryConfigs): The configuration object for the memory system,
containing initialization parameters for all submodules.
Components initialized:
- compressor (optional): Pre-compression model if pre_compress=True
- segmenter (optional): Topic segmentation model if topic_segment=True
- manager: Memory management model for metadata generation and text summarization
- text_embedder (optional): Text embedding model if index_strategy is 'embedding' or 'hybrid'
- retrieve_strategy (optional): Retrieval strategy ('context', 'embedding', or 'hybrid')
- context_retriever (optional): Context-based retriever if retrieve_strategy is 'context' or 'hybrid'
- embedding_retriever (optional): Embedding-based retriever if retrieve_strategy is 'embedding' or 'hybrid'
- graph (optional): Graph memory store if graph_mem is enabled
Note:
- Multimodal embedder initialization is currently commented out
- Graph memory initialization is conditional on graph_mem configuration
"""
if config.logging is not None:
config.logging.apply()
self.logger = get_logger("LightMemory")
self.logger.info("Initializing LightMemory with provided configuration")
self.token_stats = {
"add_memory_calls": 0,
"add_memory_prompt_tokens": 0,
"add_memory_completion_tokens": 0,
"add_memory_total_tokens": 0,
"update_calls": 0,
"update_prompt_tokens": 0,
"update_completion_tokens": 0,
"update_total_tokens": 0,
"embedding_calls": 0,
"embedding_total_tokens": 0,
"summarize_calls": 0,
"summarize_prompt_tokens": 0,
"summarize_completion_tokens": 0,
"summarize_total_tokens": 0,
"embedding_calls": 0,
"embedding_total_tokens": 0,
}
self.logger.info("Token statistics tracking initialized")
self.config = config
self.compressor = None
self.segmenter = None
if self.config.pre_compress:
self.logger.info("Initializing pre-compressor")
self.compressor = PreCompressorFactory.from_config(self.config.pre_compressor)
if self.config.topic_segment:
self.logger.info("Initializing topic segmenter")
self.segmenter = TopicSegmenterFactory.from_config(self.config.topic_segmenter, self.config.precomp_topic_shared, self.compressor)
self.senmem_buffer_manager = SenMemBufferManager(max_tokens=self.segmenter.buffer_len, tokenizer=self.segmenter.tokenizer)
self.logger.info("Initializing memory manager")
self.manager = MemoryManagerFactory.from_config(self.config.memory_manager)
self.shortmem_buffer_manager = ShortMemBufferManager(max_tokens = 512, tokenizer=getattr(self.manager, "tokenizer", self.manager.config.model))
if self.config.index_strategy == 'embedding' or self.config.index_strategy == 'hybrid':
self.logger.info("Initializing text embedder")
self.text_embedder = TextEmbedderFactory.from_config(self.config.text_embedder)
# if self.config.multimodal_embedder:
self.retrieve_strategy = self.config.retrieve_strategy
if self.retrieve_strategy in ["context", "hybrid"]:
self.logger.info("Initializing context retriever")
self.context_retriever = ContextRetrieverFactory.from_config(self.config.context_retriever)
if self.retrieve_strategy in ["embedding", "hybrid"]:
self.logger.info("Initializing embedding retriever")
self.embedding_retriever = EmbeddingRetrieverFactory.from_config(self.config.embedding_retriever)
if hasattr(self.config, 'summary_retriever') and self.config.summary_retriever is not None:
self.logger.info("Initializing summary retriever")
self.summary_retriever = EmbeddingRetrieverFactory.from_config(self.config.summary_retriever)
if self.config.graph_mem:
from .graph import GraphMem
self.logger.info("Initializing graph memory")
self.graph = GraphMem(self.config.graph_mem)
self.logger.info("LightMemory initialization completed successfully")
@classmethod
def from_config(cls, config: Dict[str,Any]):
try:
configs = BaseMemoryConfigs(**config)
except ValidationError as e:
print(f"Configuration validation error: {e}")
raise
return cls(configs)
def add_memory(
self,
messages,
METADATA_GENERATE_PROMPT: Optional[Union[str, Dict[str, str]]] = None,
*,
force_segment: bool = False,
force_extract: bool = False
):
"""
Add new memory entries from message history.
This method serves as the main pipeline for constructing new memory units from
incoming messages. It performs message normalization, optional pre-compression,
segmentation, and knowledge extraction to produce structured memory entries.
The process is as follows:
1. Normalize input messages with standardized timestamps and session tracking.
2. Optionally compress messages using the pre-defined compression model (if enabled).
3. If topic segmentation is enabled, split messages into coherent segments and add them to the sentence-level buffer.
4. Trigger memory extraction based on configured thresholds or forced flags.
5. Optionally perform metadata summarization using an external model if enabled.
6. Convert extracted results into `MemoryEntry` objects and update memory storage
(either in online or offline mode depending on configuration).
Args:
messages (dict or List[dict]): Input message(s) to process.
METADATA_GENERATE_PROMPT: Custom prompt(s) for extraction. Supports multiple formats:
- str: Legacy format for flat mode (single factual prompt)
Example: METADATA_GENERATE_PROMPT="Your extraction prompt..."
- dict: New format supporting multiple perspectives
For flat mode: {"factual": "..."}
For event mode: {"factual": "...", "relational": "..."}
- None: Use default prompts based on self.config.extraction_mode
force_segment (bool, optional): If True, forces segmentation regardless of buffer conditions.
force_extract (bool, optional): If True, forces memory extraction even if thresholds are not met.
Returns:
dict: A dictionary containing the intermediate results of the memory addition pipeline.
Typically includes:
- `"add_input_prompt"`: List of input prompts used for metadata generation (if enabled)
- `"add_output_prompt"`: Corresponding output results from metadata generation
- `"api_call_nums"`: Number of API calls made for extraction/summarization
- (In early termination cases) A segmentation result dict with keys such as
`"triggered"`, `"cut_index"`, `"boundaries"`, and `"emitted_messages"`
Notes:
- If `self.config.pre_compress` is True, messages will first be token-compressed before segmentation.
- If `self.config.topic_segment` is disabled, the function returns early with segmentation info only.
- Memory extraction results are wrapped into `MemoryEntry` objects containing timestamps,
weekdays, and extracted factual content.
- Depending on `self.config.update`, the function triggers either online or offline memory updates.
"""
extract_prompts = normalize_extraction_prompts(
prompts=METADATA_GENERATE_PROMPT,
extraction_mode=self.config.extraction_mode,
logger=self.logger
)
call_id = f"add_memory_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
self.logger.info(f"========== START {call_id} ==========")
self.logger.info(f"force_segment={force_segment}, force_extract={force_extract}")
result = {
"add_input_prompt": [],
"add_output_prompt": [],
"api_call_nums": 0
}
self.logger.debug(f"[{call_id}] Raw input type: {type(messages)}")
if isinstance(messages, list):
self.logger.debug(f"[{call_id}] Raw input sample: {json.dumps(messages)}")
normalizer = MessageNormalizer(offset_ms=500)
msgs = normalizer.normalize_messages(messages)
self.logger.debug(f"[{call_id}] Normalized messages sample: {json.dumps(msgs)}")
if self.config.pre_compress:
if hasattr(self.compressor, "tokenizer") and self.compressor.tokenizer is not None:
args = (msgs, self.compressor.tokenizer)
elif self.config.topic_segment and hasattr(self.segmenter, "tokenizer") and self.segmenter.tokenizer is not None:
args = (msgs, self.segmenter.tokenizer)
else:
args = (msgs,)
# fixed: empty 'content' in the 'messages' of 'compress(*args)'
compressed_messages = self.compressor.compress(*args)
cfg = getattr(self.compressor, "config", None)
target_rate = None
if cfg is not None:
if hasattr(cfg, 'entropy_config') and isinstance(cfg.entropy_config, dict):
target_rate = cfg.entropy_config.get('compress_rate')
elif hasattr(cfg, 'compress_config') and isinstance(cfg.compress_config, dict):
target_rate = cfg.compress_config.get('rate')
self.logger.info(f"[{call_id}] Target compression rate: {target_rate}")
self.logger.debug(f"[{call_id}] Compressed messages sample: {json.dumps(compressed_messages)}")
else:
compressed_messages = msgs
self.logger.info(f"[{call_id}] Pre-compression disabled, using normalized messages")
if not self.config.topic_segment:
# TODO:
self.logger.info(f"[{call_id}] Topic segmentation disabled, returning emitted messages")
return {
"triggered": True,
"cut_index": len(msgs),
"boundaries": [0, len(msgs)],
"emitted_messages": msgs,
"carryover_size": 0,
}
all_segments = self.senmem_buffer_manager.add_messages(compressed_messages, self.segmenter, self.text_embedder)
if force_segment:
all_segments = self.senmem_buffer_manager.cut_with_segmenter(self.segmenter, self.text_embedder, force_segment)
if not all_segments:
self.logger.debug(f"[{call_id}] No segments generated, returning empty result")
return result # TODO
self.logger.info(f"[{call_id}] Generated {len(all_segments)} segments")
self.logger.debug(f"[{call_id}] Segments sample: {json.dumps(all_segments)}")
extract_trigger_num, extract_list = self.shortmem_buffer_manager.add_segments(all_segments, self.config.messages_use, force_extract)
if extract_trigger_num == 0:
self.logger.debug(f"[{call_id}] Extraction not triggered, returning result")
return result # TODO
global GLOBAL_TOPIC_IDX
topic_id_mapping = []
for api_call_segments in extract_list:
api_call_topic_ids = []
for topic_segment in api_call_segments:
api_call_topic_ids.append(GLOBAL_TOPIC_IDX)
GLOBAL_TOPIC_IDX += 1
topic_id_mapping.append(api_call_topic_ids)
self.logger.debug(f"topic_id_mapping: {topic_id_mapping}")
self.logger.info(f"[{call_id}] Assigned global topic IDs: total={sum(len(x) for x in topic_id_mapping)}, mapping={topic_id_mapping}")
self.logger.info(f"[{call_id}] Extraction triggered {extract_trigger_num} times, extract_list length: {len(extract_list)}")
extract_list, timestamps_list, weekday_list, speaker_list, topic_id_map = assign_sequence_numbers_with_timestamps(extract_list, offset_ms=500, topic_id_mapping=topic_id_mapping)
self.logger.debug(f"[{call_id}] Extract list sample: {json.dumps(extract_list)}")
max_source_ids = [sum(1 for seg in batch for msg in seg if msg.get("role") == "user") - 1 for batch in extract_list]
self.logger.info(f"[{call_id}] Batch max_source_ids: {max_source_ids}")
if self.config.metadata_generate and self.config.text_summary:
self.logger.info(f"[{call_id}] Starting metadata generation")
extracted_results = self.manager.meta_text_extract(
extract_list=extract_list,
messages_use=self.config.messages_use,
topic_id_mapping=topic_id_mapping,
extraction_mode=self.config.extraction_mode,
custom_prompts=extract_prompts
)
# ============ API token Consumption ============
process_extraction_results(
extracted_results=extracted_results,
token_stats=self.token_stats,
result_dict=result,
call_id=call_id,
logger=self.logger
)
self.logger.info(f"[{call_id}] Metadata generation completed with {result['api_call_nums']} API calls")
memory_entries = convert_extraction_results_to_memory_entries(
extracted_results=extracted_results,
timestamps_list=timestamps_list,
weekday_list=weekday_list,
speaker_list=speaker_list,
topic_id_map=topic_id_map,
max_source_ids=max_source_ids,
logger=self.logger
)
self.logger.info(f"[{call_id}] Created {len(memory_entries)} MemoryEntry objects")
for i, mem in enumerate(memory_entries):
self.logger.debug(f"[{call_id}] MemoryEntry[{i}]: time={mem.time_stamp}, weekday={mem.weekday}, speaker_id={mem.speaker_id}, speaker_name={mem.speaker_name}, topic_id={mem.topic_id}, memory={mem.memory}")
if self.config.update == "online":
self.online_update(memory_entries)
elif self.config.update == "offline":
self.offline_update(memory_entries)
self.logger.info(
f"[{call_id}] Cumulative token stats - "
f"Total API calls: {self.token_stats['add_memory_calls']}, "
f"Total tokens: {self.token_stats['add_memory_total_tokens']}"
)
return result
def online_update(self, memory_list: List):
return None
def offline_update(self, memory_list: List, construct_update_queue_trigger: bool = False, offline_update_trigger: bool = False):
call_id = f"offline_update_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
self.logger.info(f"========== START {call_id} ==========")
self.logger.info(f"[{call_id}] Received {len(memory_list)} memory entries")
self.logger.info(f"[{call_id}] construct_update_queue_trigger={construct_update_queue_trigger}, offline_update_trigger={offline_update_trigger}")
if self.config.index_strategy in ["context", "hybrid"]:
self.logger.info(f"[{call_id}] Saving memory entries to file (strategy: {self.config.index_strategy})")
save_memory_entries(memory_list, "memory_entries.json")
if self.config.index_strategy in ["embedding", "hybrid"]:
inserted_count = 0
self.logger.info(f"[{call_id}] Starting embedding and insertion to vector database")
embeddings = self.text_embedder.embed([mem_obj.memory for mem_obj in memory_list]) if memory_list else []
batched_vectors = []
batched_payloads = []
batched_ids = []
for mem_obj, embedding_vector in zip(memory_list, embeddings):
ids = mem_obj.id
while self.embedding_retriever.exists(ids):
ids = str(uuid.uuid4())
mem_obj.id = ids
payload = {
"time_stamp": mem_obj.time_stamp,
"float_time_stamp": mem_obj.float_time_stamp,
"weekday": mem_obj.weekday,
"topic_id": mem_obj.topic_id,
"topic_summary": mem_obj.topic_summary,
"category": mem_obj.category,
"subcategory": mem_obj.subcategory,
"memory_class": mem_obj.memory_class,
"memory": mem_obj.memory,
"original_memory": mem_obj.original_memory,
"compressed_memory": mem_obj.compressed_memory,
"speaker_id": mem_obj.speaker_id,
"speaker_name": mem_obj.speaker_name,
"consolidated": mem_obj.consolidated,
}
batched_vectors.append(embedding_vector)
batched_payloads.append(payload)
batched_ids.append(ids)
inserted_count += 1
if batched_vectors:
self.embedding_retriever.insert(
vectors=batched_vectors,
payloads=batched_payloads,
ids=batched_ids,
)
self.logger.info(f"[{call_id}] Successfully inserted {inserted_count} entries to vector database")
if construct_update_queue_trigger:
self.logger.info(f"[{call_id}] Triggering update queue construction")
self.construct_update_queue_all_entries(
top_k=20,
keep_top_n=10
)
if offline_update_trigger:
self.logger.info(f"[{call_id}] Triggering offline update for all entries")
self.offline_update_all_entries(
update_sim_threshold = 0.8
)
def construct_update_queue_all_entries(self, top_k: int = 20, keep_top_n: int = 10, max_workers: int = 8):
"""
Offline update all entries in parallel using multithreading.
Each entry updates its own update_queue based on entries with earlier timestamps.
Args:
top_k (int): Number of nearest neighbors to consider for each entry.
keep_top_n (int): Number of top entries to keep in update_queue.
max_workers (int): Maximum number of threads to use.
"""
call_id = f"construct_queue_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
self.logger.info(f"========== START {call_id} ==========")
self.logger.info(f"[{call_id}] Parameters: top_k={top_k}, keep_top_n={keep_top_n}, max_workers={max_workers}")
all_entries = self.embedding_retriever.get_all()
self.logger.info(f"[{call_id}] Retrieved {len(all_entries)} entries from vector database")
if not all_entries:
self.logger.warning(f"[{call_id}] No entries found in database, skipping queue construction")
self.logger.info(f"========== END {call_id} ==========")
return
updated_count = 0
skipped_count = 0
nonempty_queue_count = 0
empty_queue_count = 0
lock = threading.Lock()
write_lock = threading.Lock()
def _update_queue_construction(entry):
nonlocal updated_count, skipped_count, nonempty_queue_count, empty_queue_count
eid = entry["id"]
payload = entry["payload"]
vec = entry.get("vector")
ts = payload.get("float_time_stamp", None)
if vec is None or ts is None:
self.logger.debug(f"[{call_id}] Skipping entry {eid}: missing vector={vec is None}, float_time_stamp={ts is None} ({ts})")
with lock:
skipped_count += 1
return
hits = self.embedding_retriever.search(
query_vector=vec,
limit=top_k,
filters={"float_time_stamp": {"lte": ts}}
)
candidates = []
for h in hits:
hid = h["id"]
if hid == eid:
continue
candidates.append({"id": hid, "score": h.get("score")})
candidates.sort(key=lambda x: x["score"], reverse=True)
update_queue = candidates[:keep_top_n]
new_payload = dict(payload)
new_payload["update_queue"] = update_queue
if update_queue:
with lock:
nonempty_queue_count += 1
self.logger.debug(f"[{call_id}] Entry {eid} update_queue length={len(update_queue)} top_candidates=" + str(update_queue[:3]))
else:
with lock:
empty_queue_count += 1
self.logger.debug(f"[{call_id}] Entry {eid} has no candidates after filtering (hits may be only itself)")
with write_lock:
self.embedding_retriever.update(vector_id=eid, vector=vec, payload=new_payload)
with lock:
updated_count += 1
self.logger.info(f"[{call_id}] Starting parallel queue construction with {max_workers} workers")
with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
executor.map(_update_queue_construction, all_entries)
self.logger.info(
f"[{call_id}] Queue construction completed: {updated_count} updated, {skipped_count} skipped, "
f"nonempty_queues={nonempty_queue_count}, empty_queues={empty_queue_count}"
)
self.logger.info(f"========== END {call_id} ==========")
def offline_update_all_entries(self, score_threshold: float = 0.9, max_workers: int = 5):
"""
Perform offline updates for all entries based on their update_queue, in parallel.
Args:
score_threshold (float): Minimum similarity score for considering update candidates.
max_workers (int): Maximum number of worker threads.
"""
call_id = f"offline_update_all_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
self.logger.info(f"========== START {call_id} ==========")
self.logger.info(f"[{call_id}] Parameters: score_threshold={score_threshold}, max_workers={max_workers}")
all_entries = self.embedding_retriever.get_all()
self.logger.info(f"[{call_id}] Retrieved {len(all_entries)} entries from vector database")
if not all_entries:
self.logger.warning(f"[{call_id}] No entries found in database, skipping offline update")
self.logger.info(f"========== END {call_id} ==========")
return
processed_count = 0
updated_count = 0
deleted_count = 0
skipped_count = 0
lock = threading.Lock()
write_lock = threading.Lock()
update_token_stats = {
"calls": 0,
"prompt_tokens": 0,
"completion_tokens": 0,
"total_tokens": 0
}
token_lock = threading.Lock()
def update_entry(entry):
nonlocal processed_count, updated_count, deleted_count, skipped_count
eid = entry["id"]
payload = entry["payload"]
candidate_sources = []
for other in all_entries:
update_queue = other["payload"].get("update_queue", [])
for candidate in update_queue:
if candidate["id"] == eid and candidate["score"] >= score_threshold:
candidate_sources.append(other)
break
if not candidate_sources:
with lock:
skipped_count += 1
return
with lock:
processed_count += 1
updated_entry = self.manager._call_update_llm(UPDATE_PROMPT, entry, candidate_sources)
if updated_entry is None:
return
# ====== token consumption ======
usage = updated_entry["usage"]
with token_lock:
update_token_stats["calls"] += 1
update_token_stats["prompt_tokens"] += usage.get("prompt_tokens", 0)
update_token_stats["completion_tokens"] += usage.get("completion_tokens", 0)
update_token_stats["total_tokens"] += usage.get("total_tokens", 0)
self.logger.debug(
f"[{call_id}] Update LLM call for {eid} - "
f"Tokens: {usage.get('total_tokens', 0)}"
)
# ==================== token consumption ====================
action = updated_entry.get("action")
if action == "delete":
with write_lock:
self.embedding_retriever.delete(eid)
with lock:
deleted_count += 1
self.logger.debug(f"[{call_id}] Deleted entry: {eid}")
elif action == "update":
new_payload = dict(payload)
new_payload["memory"] = updated_entry.get("new_memory")
vector = entry.get("vector")
with write_lock:
self.embedding_retriever.update(vector_id=eid, vector=vector, payload=new_payload)
with lock:
updated_count += 1
self.logger.debug(f"[{call_id}] Updated entry: {eid}")
self.logger.info(f"[{call_id}] Starting parallel offline update with {max_workers} workers")
with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
executor.map(update_entry, all_entries)
with lock:
self.token_stats["update_calls"] += update_token_stats["calls"]
self.token_stats["update_prompt_tokens"] += update_token_stats["prompt_tokens"]
self.token_stats["update_completion_tokens"] += update_token_stats["completion_tokens"]
self.token_stats["update_total_tokens"] += update_token_stats["total_tokens"]
self.logger.info(f"[{call_id}] Offline update completed:")
self.logger.info(f"[{call_id}] - Processed: {processed_count} entries")
self.logger.info(f"[{call_id}] - Updated: {updated_count} entries")
self.logger.info(f"[{call_id}] - Deleted: {deleted_count} entries")
self.logger.info(f"[{call_id}] - Skipped (no candidates): {skipped_count} entries")
self.logger.info(
f"[{call_id}] - Update API calls: {update_token_stats['calls']}, "
f"Total tokens: {update_token_stats['total_tokens']}"
)
self.logger.info(f"========== END {call_id} ==========")
def retrieve(self, query: str, limit: int = 10, filters: Optional[dict] = None) -> list[str]:
"""
Retrieve relevant entries and return them as formatted strings.
Args:
query (str): The natural language query string.
limit (int, optional): Number of results to return. Defaults to 10.
filters (dict, optional): Optional filters to narrow down the search. Defaults to None.
Returns:
list[str]: A list of formatted strings containing time_stamp, weekday, and memory.
"""
call_id = f"retrieve_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
self.logger.info(f"========== START {call_id} ==========")
self.logger.info(f"[{call_id}] Query: {query}")
self.logger.info(f"[{call_id}] Parameters: limit={limit}, filters={filters}")
self.logger.debug(f"[{call_id}] Generating embedding for query")
query_vector = self.text_embedder.embed(query)
self.logger.debug(f"[{call_id}] Query embedding dimension: {len(query_vector)}")
self.logger.info(f"[{call_id}] Searching vector database")
results = self.embedding_retriever.search(
query_vector=query_vector,
limit=limit,
filters=filters,
return_full=True,
)
self.logger.info(f"[{call_id}] Found {len(results)} results")
formatted_results = []
for r in results:
payload = r.get("payload", {})
time_stamp = payload.get("time_stamp", "")
weekday = payload.get("weekday", "")
memory = payload.get("memory", "")
formatted_results.append(f"{time_stamp} {weekday} {memory}")
result_string = "\n".join(formatted_results)
self.logger.info(f"[{call_id}] Formatted {len(formatted_results)} results into output string")
self.logger.debug(f"[{call_id}] Output string length: {len(result_string)} characters")
self.logger.info(f"========== END {call_id} ==========")
return result_string
def get_token_statistics(self):
embedder_stats = {"total_calls": 0, "total_tokens": None}
if hasattr(self, 'text_embedder') and hasattr(self.text_embedder, 'get_stats'):
embedder_stats = self.text_embedder.get_stats()
stats = {
"summary": {
"total_llm_calls": self.token_stats["add_memory_calls"] + self.token_stats["update_calls"] + self.token_stats["summarize_calls"],
"total_llm_tokens": self.token_stats["add_memory_total_tokens"] + self.token_stats["update_total_tokens"] + self.token_stats["summarize_total_tokens"],
"total_embedding_calls": embedder_stats["total_calls"],
"total_embedding_tokens": embedder_stats["total_tokens"],
},
"llm": {
"add_memory": {
"calls": self.token_stats["add_memory_calls"],
"prompt_tokens": self.token_stats["add_memory_prompt_tokens"],
"completion_tokens": self.token_stats["add_memory_completion_tokens"],
"total_tokens": self.token_stats["add_memory_total_tokens"],
},
"update": {
"calls": self.token_stats["update_calls"],
"prompt_tokens": self.token_stats["update_prompt_tokens"],
"completion_tokens": self.token_stats["update_completion_tokens"],
"total_tokens": self.token_stats["update_total_tokens"],
},
"summarize": {
"calls": self.token_stats["summarize_calls"],
"prompt_tokens": self.token_stats["summarize_prompt_tokens"],
"completion_tokens": self.token_stats["summarize_completion_tokens"],
"total_tokens": self.token_stats["summarize_total_tokens"],
},
},
"embedding": {
"total_calls": embedder_stats["total_calls"],
"total_tokens": embedder_stats["total_tokens"],
"note": "Includes topic segmentation + memory indexing. Local models show None for tokens."
}
}
return stats
def summarize(
self,
SUMMARY_PROMPT: Optional[str] = None,
*,
time_window: int = 3600,
process_all: bool = False,
enable_cross_event: bool = True,
retrieval_scope: Literal["global", "historical"] = "global",
top_k_seeds: int = 15,
) -> Dict:
from lightmem.memory.utils import (
initialize_time_pointer,
get_window_entries,
mark_entries_and_get_next_time,
check_has_more_entries,
retrieve_supplementary_entries,
format_entries_for_prompt,
call_summary_llm,
store_summary,
build_summary_item,
build_single_result,
build_batch_result,
build_empty_result
)
global GLOBAL_LAST_SUMMARY_TIME
call_id = f"summarize_{'all' if process_all else 'once'}_{datetime.now().strftime('%Y%m%d_%H%M%S_%f')}"
self.logger.info(f"========== START {call_id} ==========")
if not self.summary_retriever:
raise ValueError("Summarization not enabled. Set 'summary_collection_name' in config.")
summaries = [] if process_all else None
total_entries = 0 if process_all else None
iteration = 0
while True:
iteration += 1
self.logger.info(f"[{call_id}] Iteration {iteration}")
if GLOBAL_LAST_SUMMARY_TIME is None:
GLOBAL_LAST_SUMMARY_TIME = initialize_time_pointer(
retriever=self.embedding_retriever,
call_id=call_id,
logger=self.logger
)
if GLOBAL_LAST_SUMMARY_TIME is None:
return build_empty_result(process_all)
Cbuf, has_more, new_time = get_window_entries(
retriever=self.embedding_retriever,
current_time=GLOBAL_LAST_SUMMARY_TIME,
time_window=time_window,
call_id=call_id,
logger=self.logger
)
if Cbuf is None:
if new_time is not None:
GLOBAL_LAST_SUMMARY_TIME = new_time
if process_all:
if has_more:
continue
else:
break
else:
return build_empty_result(process_all, has_more=has_more)
self.logger.info(f"[{call_id}] Processing {len(Cbuf)} entries")
Sk = []
if enable_cross_event:
retrieval_filters = None
if retrieval_scope == "historical":
retrieval_filters = {
"float_time_stamp": {"lt": Cbuf[0]["payload"]["float_time_stamp"]}
}
Sk = retrieve_supplementary_entries(
buffer_entries=Cbuf,
retriever=self.embedding_retriever,
text_embedder=self.text_embedder,
top_k=top_k_seeds,
retrieval_scope=retrieval_scope,
additional_filters=retrieval_filters,
logger=self.logger
)
self.logger.debug(f"[{call_id}] Retrieved {len(Sk)} seeds")
has_entry_type = any(e["payload"].get("entry_type") for e in Cbuf)
buffer_text = format_entries_for_prompt(Cbuf, include_type_tag=has_entry_type)
supplementary_text = format_entries_for_prompt(Sk, include_type_tag=has_entry_type)
time_range_str = f"{Cbuf[0]['payload']['time_stamp']} - {Cbuf[-1]['payload']['time_stamp']}"
speakers = list(set(
e["payload"].get("speaker_name") or e["payload"].get("speaker_id") or "?"
for e in Cbuf
))
summary_text = call_summary_llm(
manager=self.manager,
buffer_text=buffer_text,
supplementary_text=supplementary_text,
time_range=time_range_str,
speakers=speakers,
custom_prompt=SUMMARY_PROMPT,
token_stats=self.token_stats,
logger=self.logger
)
self.logger.debug(f"[{call_id}] Generated {len(summary_text)} chars")
summary_id = store_summary(
summary_text=summary_text,
buffer_entries=Cbuf,
seed_entries=Sk,
summary_retriever=self.summary_retriever,
text_embedder=self.text_embedder,
logger=self.logger
)
GLOBAL_LAST_SUMMARY_TIME = mark_entries_and_get_next_time(
retriever=self.embedding_retriever,
entries=Cbuf,
call_id=call_id,
logger=self.logger
)
has_more = check_has_more_entries(
retriever=self.embedding_retriever,
current_time=GLOBAL_LAST_SUMMARY_TIME
)
if process_all:
summaries.append(build_summary_item(summary_text, summary_id, Cbuf, Sk))
total_entries += len(Cbuf)
if not has_more:
break
else:
result = build_single_result(summary_text, summary_id, Cbuf, Sk, has_more)
self.logger.info(f"========== END {call_id} ==========")
return result
result = build_batch_result(summaries, total_entries, call_id, self.logger)
self.logger.info(f"========== END {call_id} ==========")
return result
|