File size: 25,495 Bytes
22087d0 | 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 | """
Thread Semantic Search Service
Provides vector-based semantic search for threads using OpenAI embeddings and Supabase pgvector.
"""
import os
import json
import uuid
from typing import List, Optional, Tuple
from dotenv import load_dotenv
load_dotenv()
from core.utils.logger import logger
# Chunk configuration
CHUNK_SIZE = 1200
CHUNK_OVERLAP = 100
class RecursiveCharacterTextSplitter:
"""
Recursively splits text using a hierarchy of separators.
Prioritizes keeping User:/Assistant: message boundaries intact.
"""
def __init__(
self,
chunk_size: int = 1200,
chunk_overlap: int = 100,
separators: List[str] = None
):
self.chunk_size = chunk_size
self.chunk_overlap = chunk_overlap
self.separators = separators or [
"\n\nUser: ",
"\n\nAssistant: ",
"\n\n",
"\n",
". ",
"? ",
"! ",
"; ",
", ",
" ",
""
]
def split_text(self, text: str) -> List[str]:
"""Split text into chunks respecting message boundaries."""
if not text:
return []
if len(text) <= self.chunk_size:
return [text.strip()] if text.strip() else []
return self._split_recursive(text, self.separators)
def _split_recursive(self, text: str, separators: List[str]) -> List[str]:
"""Recursively split text using the separator hierarchy."""
final_chunks = []
separator = separators[-1]
new_separators = []
for i, sep in enumerate(separators):
if sep == "":
separator = sep
new_separators = []
break
if sep in text:
separator = sep
new_separators = separators[i + 1:]
break
if separator:
splits = text.split(separator)
splits = [s + separator if i < len(splits) - 1 else s
for i, s in enumerate(splits)]
else:
splits = list(text)
current_chunk = ""
for split in splits:
split = split.strip() if not separator else split
if not split:
continue
if len(current_chunk) + len(split) > self.chunk_size:
if current_chunk.strip():
final_chunks.append(current_chunk.strip())
if len(split) > self.chunk_size:
if new_separators:
sub_chunks = self._split_recursive(split, new_separators)
final_chunks.extend(sub_chunks)
current_chunk = ""
else:
for i in range(0, len(split), self.chunk_size - self.chunk_overlap):
chunk = split[i:i + self.chunk_size].strip()
if chunk:
final_chunks.append(chunk)
current_chunk = ""
else:
if final_chunks and self.chunk_overlap > 0:
overlap_text = final_chunks[-1][-self.chunk_overlap:]
current_chunk = overlap_text + " " + split
else:
current_chunk = split
else:
current_chunk += split
if current_chunk.strip():
final_chunks.append(current_chunk.strip())
return [c for c in final_chunks if c]
# Global splitter instance
text_splitter = RecursiveCharacterTextSplitter(
chunk_size=CHUNK_SIZE,
chunk_overlap=CHUNK_OVERLAP
)
class SearchResult:
"""A single search result with thread_id, relevance score, and text preview."""
def __init__(self, thread_id: str, score: float, text_preview: str = ""):
self.thread_id = thread_id
self.score = score
self.text_preview = text_preview
class ThreadSearchService:
"""
Singleton service for thread semantic search using Supabase pgvector.
Handles embedding creation and similarity search for threads.
"""
_instance: Optional["ThreadSearchService"] = None
_initialized: bool = False
# Configuration
EMBEDDING_MODEL = "text-embedding-3-small"
EMBEDDING_DIMENSION = 1536
def __new__(cls) -> "ThreadSearchService":
if cls._instance is None:
cls._instance = super().__new__(cls)
return cls._instance
def __init__(self):
if self._initialized:
return
self._openai_client = None
self._is_configured = False
# Check if OpenAI API key is available
self._openai_api_key = os.getenv("OPENAI_API_KEY")
if self._openai_api_key:
self._is_configured = True
logger.info("[ThreadSearch] Service configured with OpenAI + Supabase pgvector")
else:
logger.warning("[ThreadSearch] Service not configured. Missing: OPENAI_API_KEY")
self._initialized = True
@property
def is_configured(self) -> bool:
"""Check if the service is properly configured."""
return self._is_configured
def _get_openai_client(self):
"""Lazily initialize AsyncOpenAI client."""
if self._openai_client is None and self._is_configured:
try:
from openai import AsyncOpenAI
self._openai_client = AsyncOpenAI(api_key=self._openai_api_key)
logger.debug("[ThreadSearch] AsyncOpenAI client initialized")
except Exception as e:
logger.error(f"[ThreadSearch] Failed to initialize OpenAI client: {e}")
self._openai_client = None
return self._openai_client
async def _create_embedding(self, text: str) -> Optional[List[float]]:
"""Create an embedding vector for the given text (async, non-blocking)."""
openai = self._get_openai_client()
if not openai:
logger.error("[ThreadSearch] _create_embedding: No OpenAI client available")
return None
try:
max_chars = 30000
if len(text) > max_chars:
text = text[:max_chars]
response = await openai.embeddings.create(
input=text,
model=self.EMBEDDING_MODEL
)
return response.data[0].embedding
except Exception as e:
logger.error(f"[ThreadSearch] Failed to create embedding: {type(e).__name__}: {e}")
return None
def _extract_snippet(self, text: str, query: str, snippet_length: int = 120) -> Tuple[str, List[str]]:
"""
Extract a snippet from text centered around matching keywords.
Prioritizes areas with the most keyword matches.
Returns:
Tuple of (snippet with "..." if truncated, list of matched keywords)
"""
if not text or not query:
snippet = text[:snippet_length] if text else ""
return (snippet + "..." if text and len(text) > snippet_length else snippet, [])
text_lower = text.lower()
query_words = [w for w in query.lower().split() if len(w) > 2]
if not query_words:
snippet = text[:snippet_length]
return (snippet + "..." if len(text) > snippet_length else snippet, [])
# Find all positions of each keyword
keyword_positions = []
matched_keywords = []
for word in query_words:
pos = 0
while True:
pos = text_lower.find(word, pos)
if pos == -1:
break
keyword_positions.append((pos, word))
if word not in matched_keywords:
matched_keywords.append(word)
pos += 1
# If no matches, return start of text
if not keyword_positions:
snippet = text[:snippet_length]
return (snippet + "..." if len(text) > snippet_length else snippet, [])
# Sort by position
keyword_positions.sort(key=lambda x: x[0])
# Find the best window - one with most keyword matches
best_count = 0
best_pos = keyword_positions[0][0]
for pos, _ in keyword_positions:
window_start = max(0, pos - 20)
window_end = window_start + snippet_length
count = sum(1 for p, _ in keyword_positions if window_start <= p < window_end)
if count > best_count:
best_count = count
best_pos = pos
# Calculate window centered on best position
half_len = snippet_length // 2
start = max(0, best_pos - half_len)
end = min(len(text), start + snippet_length)
# Adjust to not cut words
if start > 0:
space_pos = text.find(' ', start)
if space_pos != -1 and space_pos < start + 20:
start = space_pos + 1
if end < len(text):
space_pos = text.rfind(' ', start, end)
if space_pos != -1 and space_pos > end - 20:
end = space_pos
snippet = text[start:end].strip()
# Add ellipsis
prefix = "..." if start > 0 else ""
suffix = "..." if end < len(text) else ""
return (f"{prefix}{snippet}{suffix}", matched_keywords)
async def embed_thread(
self,
thread_id: str,
account_id: str,
content: str,
project_name: str = "",
thread_name: str = ""
) -> bool:
"""
Create or update embeddings for a thread in Supabase.
Uses recursive character splitting to create multiple chunks.
Args:
thread_id: The thread's unique identifier
account_id: The account that owns the thread
content: The text content to embed (full conversation)
project_name: Name of the project (for context)
thread_name: Name of the thread (for context)
Returns:
True if at least one embedding was created successfully, False otherwise
"""
if not self._is_configured:
logger.debug("[ThreadSearch] Service not configured, skipping embedding")
return False
# Build header with project/thread context
header = ""
if project_name:
header = f"Project: {project_name}"
if thread_name and thread_name != project_name:
header += f" | Thread: {thread_name}" if header else f"Thread: {thread_name}"
if header and content:
text_to_embed = f"{header}\n\n{content}"
elif content:
text_to_embed = content
else:
text_to_embed = header or ""
if not text_to_embed.strip():
logger.warning(f"[ThreadSearch] No content to embed for thread {thread_id}")
return False
# Split into chunks using recursive splitter
chunks = text_splitter.split_text(text_to_embed)
try:
from core.services.db import execute_mutate
# Delete existing chunks for this thread first
await execute_mutate(
"DELETE FROM public.documents WHERE account_id = :account_id AND thread_id = :thread_id",
{"account_id": account_id, "thread_id": thread_id}
)
# Insert each chunk
success_count = 0
for chunk_content in chunks:
embedding = await self._create_embedding(chunk_content)
if not embedding:
continue
embedding_str = json.dumps(embedding)
sql = """
INSERT INTO public.documents (chunk_id, account_id, thread_id, chunk_content, embedding, last_updated_at)
VALUES (:chunk_id, :account_id, :thread_id, :chunk_content, CAST(:embedding AS vector), NOW())
"""
await execute_mutate(sql, {
"chunk_id": str(uuid.uuid4()),
"account_id": account_id,
"thread_id": thread_id,
"chunk_content": chunk_content[:1200],
"embedding": embedding_str,
})
success_count += 1
return success_count > 0
except Exception as e:
logger.error(f"[ThreadSearch] Embedding FAILED for thread {thread_id}: {e}")
return False
async def search(
self,
query: str,
account_id: str,
limit: int = 10
) -> List[SearchResult]:
"""
Search for threads semantically matching the query using pgvector.
Args:
query: The search query text
account_id: Filter results to this account only (not used currently)
limit: Maximum number of results to return
Returns:
List of SearchResult objects with thread_id and relevance score
"""
if not self._is_configured:
logger.debug("[ThreadSearch] Service not configured, returning empty results")
return []
if not query or not query.strip():
return []
logger.info(f"[ThreadSearch] Searching: \"{query}\"")
# Create embedding for the query
query_embedding = await self._create_embedding(query)
if not query_embedding:
logger.error(f"[ThreadSearch] OpenAI embedding failed for query: \"{query}\"")
return []
try:
from core.services.db import execute
# Use pgvector cosine similarity search
# 1 - (embedding <=> query_embedding) converts distance to similarity score
# Use CAST instead of :: to avoid parameter parsing issues
# sql = """
# SELECT
# thread_id,
# user_id,
# chunk_content,
# 1 - (embedding <=> CAST(:query_embedding AS vector)) as similarity
# FROM public.documents
# WHERE embedding IS NOT NULL
# AND thread_id IS NOT NULL
# ORDER BY embedding <=> CAST(:query_embedding AS vector)
# LIMIT :limit
# """
sql = """
SELECT
thread_id,
account_id,
chunk_content,
1 - (embedding <=> CAST(:query_embedding AS vector)) AS similarity
FROM public.documents
WHERE embedding IS NOT NULL
AND thread_id IS NOT NULL
AND account_id = :account_id -- filter by tenant
ORDER BY embedding <=> CAST(:query_embedding AS vector) -- closest first
LIMIT :limit;
"""
rows = await execute(sql, {
"query_embedding": json.dumps(query_embedding),
"account_id": account_id,
"limit": limit,
})
# Convert to SearchResult objects with extracted snippets
results = []
for row in rows:
thread_id = row.get("thread_id")
if thread_id:
# Convert UUID to string if needed
thread_id = str(thread_id)
full_text = row.get("chunk_content", "")
similarity = row.get("similarity", 0)
snippet, matched = self._extract_snippet(full_text, query)
results.append(SearchResult(
thread_id=thread_id,
score=float(similarity) if similarity else 0.0,
text_preview=snippet
))
if matched:
logger.debug(f"[ThreadSearch] Thread {thread_id[:8]}... matched: {matched}")
# Log results
if results:
top = results[0]
logger.info(f"[ThreadSearch] Supabase OK: {len(results)} results, top={top.score:.3f}, preview=\"{top.text_preview[:80]}\"")
else:
logger.info(f"[ThreadSearch] Supabase OK: 0 results")
return results
except Exception as e:
logger.error(f"[ThreadSearch] Supabase search FAILED: {e}")
return []
async def delete_thread_embedding(self, thread_id: str) -> bool:
"""
Delete the embedding for a thread.
Args:
thread_id: The thread's unique identifier
Returns:
True if deletion was successful or embedding didn't exist, False on error
"""
if not self._is_configured:
return True
try:
from core.services.db import execute_mutate
sql = "DELETE FROM public.documents WHERE thread_id = :thread_id"
await execute_mutate(sql, {"thread_id": thread_id})
logger.info(f"[ThreadSearch] Deleted embedding for thread {thread_id}")
return True
except Exception as e:
logger.error(f"[ThreadSearch] Failed to delete embedding for thread {thread_id}: {e}")
return False
# Global singleton instance
_thread_search_service: Optional[ThreadSearchService] = None
def get_thread_search_service() -> ThreadSearchService:
"""Get the singleton ThreadSearchService instance."""
global _thread_search_service
if _thread_search_service is None:
_thread_search_service = ThreadSearchService()
return _thread_search_service
# Convenience functions for simpler usage
async def embed_thread(
thread_id: str,
account_id: str,
content: str,
project_name: str = "",
thread_name: str = ""
) -> bool:
"""Embed a thread for semantic search."""
service = get_thread_search_service()
return await service.embed_thread(thread_id, account_id, content, project_name, thread_name)
async def search_threads(
query: str,
account_id: str,
limit: int = 10
) -> List[SearchResult]:
"""Search threads semantically."""
service = get_thread_search_service()
return await service.search(query, account_id, limit)
async def delete_thread_embedding(thread_id: str) -> bool:
"""Delete embedding for a thread."""
service = get_thread_search_service()
return await service.delete_thread_embedding(thread_id)
async def get_threads_needing_embedding(account_id: str, thread_ids: List[str]) -> set:
"""
Return set of thread_ids that need embedding (no embedding or stale embedding).
A thread needs embedding if:
1. It has no embeddings in the documents table, OR
2. It has messages newer than its latest embedding
Filters by account_id first for efficient lookups on the documents table.
Args:
account_id: The account that owns the threads
thread_ids: List of thread IDs to check
Returns:
Set of thread_ids that need (re-)embedding
"""
if not thread_ids:
return set()
try:
from core.services.db import execute
# Find threads that either:
# 1. Have no embeddings at all, OR
# 2. Have messages created after the last embedding update
# Filter documents by account_id first for index efficiency
sql = """
WITH thread_embedding_status AS (
SELECT
t.thread_id,
MAX(d.last_updated_at) as last_embedded_at,
MAX(m.updated_at) as last_message_at
FROM unnest(CAST(:thread_ids AS uuid[])) AS t(thread_id)
LEFT JOIN public.documents d
ON d.account_id = :account_id
AND d.thread_id = t.thread_id
LEFT JOIN public.messages m
ON m.thread_id = t.thread_id
AND m.type IN ('user', 'assistant')
GROUP BY t.thread_id
)
SELECT thread_id
FROM thread_embedding_status
WHERE last_embedded_at IS NULL -- Never embedded
OR last_message_at > last_embedded_at -- Has newer messages
"""
rows = await execute(sql, {
"account_id": account_id,
"thread_ids": thread_ids,
})
result = {str(row['thread_id']) for row in rows} if rows else set()
logger.info(f"[ThreadSearch] Account {account_id[:8]}...: {len(result)}/{len(thread_ids)} threads need embedding")
return result
except Exception as e:
logger.error(f"[ThreadSearch] Failed to check threads needing embedding: {e}")
# On error, return all thread_ids to be safe (will re-embed everything)
return set(thread_ids)
async def embed_thread_with_messages(thread_id: str, account_id: str) -> bool:
"""
Fetch all thread messages and embed the full conversation.
Args:
thread_id: The thread's unique identifier
account_id: The account that owns the thread
Returns:
True if embedding was successful, False otherwise
"""
service = get_thread_search_service()
if not service.is_configured:
return False
try:
from core.services.db import execute
# Fetch thread info
thread_sql = """
SELECT
t.thread_id,
t.name as thread_name,
p.name as project_name
FROM threads t
LEFT JOIN projects p ON t.project_id = p.project_id
WHERE t.thread_id = :thread_id
"""
thread_rows = await execute(thread_sql, {"thread_id": thread_id})
if not thread_rows:
logger.warning(f"[ThreadSearch] Thread {thread_id} not found for embedding")
return False
thread = thread_rows[0]
project_name = thread.get("project_name") or ""
thread_name = thread.get("thread_name") or ""
# Fetch ALL messages in the thread (user and assistant only)
messages_sql = """
SELECT
type,
content->>'content' as text_content
FROM messages
WHERE thread_id = :thread_id
AND type IN ('user', 'assistant')
AND content->>'content' IS NOT NULL
AND content->>'content' != ''
ORDER BY created_at ASC
"""
message_rows = await execute(messages_sql, {"thread_id": thread_id})
# Build conversation content
conversation_parts = []
for msg in message_rows or []:
msg_type = msg.get("type", "")
text = msg.get("text_content", "")
if text:
if msg_type == "user":
conversation_parts.append(f"User: {text}")
elif msg_type == "assistant":
conversation_parts.append(f"Assistant: {text}")
full_content = "\n\n".join(conversation_parts)
# Skip if no content to embed
if not full_content and not project_name and not thread_name:
logger.debug(f"[ThreadSearch] Thread {thread_id} has no content to embed")
return True # Not an error, just nothing to embed
success = await embed_thread(
thread_id=thread_id,
account_id=account_id,
content=full_content,
project_name=project_name,
thread_name=thread_name
)
return success
except Exception as e:
logger.error(f"[ThreadSearch] Failed to embed thread {thread_id}: {e}")
return False
async def embed_unembedded_threads(account_id: str, threads: List[dict]):
"""
Check which threads need embedding and embed them in parallel.
Runs in background, doesn't block the response.
A thread needs embedding if:
1. It has never been embedded, OR
2. It has new messages since the last embedding
Args:
account_id: The account that owns the threads
threads: List of thread dicts (must have 'thread_id' key)
"""
import asyncio
service = get_thread_search_service()
if not service.is_configured:
logger.debug("[ThreadSearch] Service not configured, skipping background embedding")
return
try:
# Get thread_ids from the list (str() to handle UUID objects from DB)
thread_ids = [str(t.get('thread_id')) for t in threads if t.get('thread_id')]
if not thread_ids:
return
# Check which threads need embedding (no embedding or stale)
needs_embedding_ids = await get_threads_needing_embedding(account_id, thread_ids)
if not needs_embedding_ids:
return
# Filter to threads needing embedding (str() to handle UUID objects)
to_embed = [t for t in threads if str(t.get('thread_id')) in needs_embedding_ids]
# Embed in parallel with semaphore to limit concurrency
semaphore = asyncio.Semaphore(5)
async def embed_one(thread: dict):
async with semaphore:
await embed_thread_with_messages(
thread_id=str(thread['thread_id']),
account_id=account_id
)
# Use gather with return_exceptions to not fail on individual errors
await asyncio.gather(
*[embed_one(t) for t in to_embed],
return_exceptions=True
)
logger.info(f"[ThreadSearch] Background embedding complete for {len(to_embed)} threads")
except Exception as e:
logger.error(f"[ThreadSearch] Background embedding failed: {e}")
|