File size: 32,563 Bytes
1525e1c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 | import logging
logger = logging.getLogger(__name__)
"""
Retrieval tools for the agentic system
Gap #1: HybridSearchTool β BM25 + Vector with RRF fusion
Gap #2: CommunitySummaryTool β LazyGraphRAG community summaries
Gap #6: Scoring utilities for Graph-of-Thought parallel exploration
"""
from typing import List, Dict, Any, Optional
import json
import asyncio
import hashlib
from ..core.neo4j_store import Neo4jStore
from ..core.llm_factory import UnifiedLLMProvider
from ..core.models import OntologySchema
from ..config import settings
# ββ Gap #1: Hybrid Search Tool (BM25 + Vector + RRF) βββββββββββββββββββββββ
class HybridSearchTool:
"""
Combines dense (vector) and sparse (BM25) retrieval via
Reciprocal Rank Fusion (RRF) β the 2025 production standard.
"""
def __init__(self, store: Neo4jStore, llm: UnifiedLLMProvider):
self.store = store
self.llm = llm
self.name = "hybrid_search"
self.description = (
"Hybrid BM25 + vector search with RRF fusion. "
"Best for most queries β handles both semantic and keyword matching."
)
async def run(
self,
query: str,
k: int = None,
filter: Optional[Dict] = None,
document_id: Optional[str] = None,
tenant_id: Optional[str] = None
) -> List[Dict[str, Any]]:
"""
Run both BM25 and vector search in parallel, then fuse with RRF.
Args:
query: Natural language query
k: Final number of results after fusion
filter: Optional metadata filters
document_id: Optional document scope
Returns:
RRF-fused results, highest-scored first
"""
k = k or settings.default_top_k
search_k = k * 3 # over-fetch for fusion
# Run both searches in parallel
query_embedding = await self.llm.embed(query)
bm25_task = self.store.bm25_search(query, k=search_k, document_id=document_id, tenant_id=tenant_id)
vector_task = self.store.search(
query_vector=query_embedding,
k=search_k,
filter=filter or ({"document_id": document_id} if document_id else None),
tenant_id=tenant_id
)
bm25_results, vector_results = await asyncio.gather(
bm25_task, vector_task, return_exceptions=True
)
if isinstance(bm25_results, Exception):
bm25_results = []
if isinstance(vector_results, Exception):
vector_results = []
# RRF Fusion
fused = self._rrf_fuse(
list_a=vector_results,
list_b=bm25_results,
rrf_k=settings.rrf_k
)
# Build final result list (top k)
doc_lookup: Dict[str, Dict] = {}
for r in vector_results + bm25_results:
if r.get("id") and r["id"] not in doc_lookup:
doc_lookup[r["id"]] = r
final = []
for chunk_id, score in fused[:k]:
doc = doc_lookup.get(chunk_id, {})
doc["hybrid_score"] = score
doc["retrieval_method"] = "hybrid_rrf"
final.append(doc)
return final
def _rrf_fuse(
self,
list_a: List[Dict],
list_b: List[Dict],
rrf_k: int = 60
) -> List[tuple]:
"""
Reciprocal Rank Fusion.
Returns sorted list of (id, score) tuples.
Non-parametric β no tuning needed.
"""
scores: Dict[str, float] = {}
for rank, doc in enumerate(list_a):
doc_id = doc.get("id", "")
if doc_id:
scores[doc_id] = scores.get(doc_id, 0.0) + 1.0 / (rrf_k + rank + 1)
for rank, doc in enumerate(list_b):
doc_id = doc.get("id", "")
if doc_id:
scores[doc_id] = scores.get(doc_id, 0.0) + 1.0 / (rrf_k + rank + 1)
return sorted(scores.items(), key=lambda x: x[1], reverse=True)
# ββ Legacy vector-only tool (kept for fallback/GoT) βββββββββββββββββββββββββ
class VectorSearchTool:
"""Vector similarity search tool β pure dense retrieval"""
def __init__(self, store: Neo4jStore, llm: UnifiedLLMProvider):
self.store = store
self.llm = llm
self.name = "vector_search"
self.description = "Semantic similarity search. Good for conceptual/thematic queries."
async def run(
self,
query: str,
k: int = None,
filter: Optional[Dict] = None
) -> List[Dict[str, Any]]:
k = k or settings.default_top_k
query_embedding = await self.llm.embed(query)
results = await self.store.search(
query_vector=query_embedding,
k=k,
filter=filter
)
for r in results:
r["retrieval_method"] = "vector"
return results
# ββ Gap #2: Community Summary Tool (LazyGraphRAG) βββββββββββββββββββββββββββ
class CommunitySummaryTool:
"""
LazyGraphRAG-style community summarization.
Detects entity clusters, generates LLM summaries at query time.
Results are cached in Redis with configurable TTL.
"""
def __init__(self, store: Neo4jStore, llm: UnifiedLLMProvider):
self.store = store
self.llm = llm
self.name = "community_search"
self.description = (
"High-level thematic search using entity communities. "
"Best for: 'What are the main themes?', 'Summarize the landscape', "
"'What is the overall picture?'"
)
self._redis = None
async def _get_redis(self):
"""Lazily initialize Redis connection for caching"""
if self._redis is None:
try:
import redis.asyncio as aioredis
self._redis = aioredis.from_url(settings.redis_url)
except Exception:
self._redis = None
return self._redis
async def run(
self,
query: str,
k: int = None,
tenant_id: Optional[str] = None
) -> List[Dict[str, Any]]:
"""
1. Find relevant entities via hybrid search
2. Group by community_id
3. Generate LLM summary for top communities (cached)
4. Return community summaries as context chunks
Args:
query: Natural language query (holistic/thematic)
k: Number of community summaries to return
Returns:
List of community summaries as context dicts
"""
k = k or 3 # Top-3 communities
# Step 1: Find relevant entities
entity_names = await self._find_relevant_entities(query, limit=20, tenant_id=tenant_id)
if not entity_names:
return []
# Step 2: Get community groupings
communities = await self.store.get_communities(entity_names, tenant_id=tenant_id)
if not communities:
return []
# Step 3: Generate summaries for top communities
results = []
for community_id, entities in list(communities.items())[:k]:
summary = await self._get_community_summary(community_id, entities, query)
if summary:
results.append({
"id": f"community_{community_id}",
"text": summary,
"community_id": community_id,
"entity_count": len(entities),
"retrieval_method": "community_summary",
"score": 0.85 # Community scores are high-confidence thematic results
})
return results
async def _find_relevant_entities(self, query: str, limit: int = 20, tenant_id: Optional[str] = None) -> List[str]:
"""Find entity names most relevant to the query via BM25"""
try:
tenant_filter_node = "WHERE node.tenant_id = $tenant_id" if tenant_id else ""
tenant_filter_entity = "{tenant_id: $tenant_id}" if tenant_id else ""
cypher = f"""
CALL db.index.fulltext.queryNodes('chunk_text_index', $query)
YIELD node, score
{tenant_filter_node}
MATCH (node)-[:MENTIONS]->(e:Entity {tenant_filter_entity})
RETURN DISTINCT e.name as name
LIMIT $limit
"""
params = {"query": query, "limit": limit}
if tenant_id:
params["tenant_id"] = tenant_id
rows = await self.store.execute_query(cypher, params)
return [r["name"] for r in rows if r.get("name")]
except Exception:
return []
async def _get_community_summary(
self,
community_id: int,
entities: List[Dict],
query: str
) -> Optional[str]:
"""Generate or fetch cached LLM summary for a community"""
# Build cache key from community entities
entity_names = sorted([e.get("name", "") for e in entities])
cache_key = f"community_summary:{hashlib.md5(':'.join(entity_names).encode()).hexdigest()}"
# Check Redis cache
redis = await self._get_redis()
if redis:
try:
cached = await redis.get(cache_key)
if cached:
return cached.decode("utf-8")
except Exception:
pass
# Generate summary with LLM
entity_descriptions = []
for e in entities[:settings.max_community_entities]:
desc = f"- {e.get('name', 'Unknown')} ({e.get('type', 'Entity')})"
entity_descriptions.append(desc)
prompt = f"""
You are analyzing a community of interconnected entities from a knowledge graph.
Community #{community_id} contains {len(entities)} entities:
{chr(10).join(entity_descriptions)}
User's question: "{query}"
Generate a focused 2-3 sentence summary of what this community represents and
how it relates to the question. Focus on themes, patterns, and key relationships.
Be specific and factual. Do not hallucinate relationships not implied by the entity names.
"""
try:
summary = await self.llm.complete(prompt, temperature=0.2)
summary = summary.strip()
# Cache the result
if redis and summary:
try:
await redis.setex(
cache_key,
settings.community_summary_cache_ttl,
summary.encode("utf-8")
)
except Exception:
pass
return summary
except Exception as e:
logger.info(f"Community summary generation error: {e}")
return None
# ββ Graph Traversal Tool βββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GraphTraversalTool:
"""Graph traversal and path finding tool"""
def __init__(self, store: Neo4jStore, llm: UnifiedLLMProvider):
self.store = store
self.llm = llm
self.name = "graph_traversal"
self.description = "Traverse knowledge graph to find relationships and paths between entities"
async def run(
self,
query: str,
source_entity: Optional[str] = None,
target_entity: Optional[str] = None,
depth: int = None,
tenant_id: Optional[str] = None
) -> List[Dict[str, Any]]:
depth = depth or settings.graph_max_depth
if not source_entity or not target_entity:
entities = await self._extract_entities_from_query(query)
if len(entities) >= 2:
source_entity = source_entity or entities[0]
target_entity = target_entity or entities[1]
elif len(entities) == 1:
source_entity = source_entity or entities[0]
if source_entity and target_entity:
results = await self.store.find_path(source_entity, target_entity, depth, tenant_id=tenant_id)
elif source_entity:
results = await self.store.get_neighbors(source_entity, depth, tenant_id=tenant_id)
else:
results = []
for r in results:
r["retrieval_method"] = "graph_traversal"
return results
async def _extract_entities_from_query(self, query: str) -> List[str]:
"""Extract entity names from natural language query"""
prompt = f"""
Extract entity names from this query:
"{query}"
Return only a JSON list of entity names: ["Entity1", "Entity2", ...]
"""
response = await self.llm.complete(prompt, temperature=0.1)
try:
cleaned = response.strip()
if "```json" in cleaned:
cleaned = cleaned.split("```json")[1].split("```")[0]
elif "```" in cleaned:
cleaned = cleaned.split("```")[1].split("```")[0]
cleaned = cleaned.strip()
entities = json.loads(cleaned)
return entities if isinstance(entities, list) else []
except Exception:
return []
# ββ Cypher Generation Tool βββββββββββββββββββββββββββββββββββββββββββββββββββ
class CypherGenerationTool:
"""
Text-to-Cypher tool with hallucination guards.
Generates Cypher queries from natural language.
"""
def __init__(
self,
store: Neo4jStore,
llm: UnifiedLLMProvider,
ontology: Optional[OntologySchema] = None
):
self.store = store
self.llm = llm
self.ontology = ontology
self.name = "cypher_query"
self.description = "Generate and execute Cypher queries for complex structured graph queries"
async def run(self, query: str, tenant_id: Optional[str] = None) -> List[Dict[str, Any]]:
if tenant_id and tenant_id != settings.default_tenant_id:
logger.warning("CypherGenerationTool disabled in multi-tenant environment to prevent data leaks.")
return []
cypher = await self._generate_cypher(query)
if not cypher:
return []
if not self._validate_cypher(cypher):
cypher = await self._correct_cypher(cypher, query)
if not self._validate_cypher(cypher):
return []
try:
results = await self.store.execute_query(cypher)
for r in results:
r["retrieval_method"] = "cypher"
return results
except Exception as e:
logger.info(f"Cypher execution error: {e}")
cypher = await self._correct_cypher_with_error(cypher, query, str(e))
try:
results = await self.store.execute_query(cypher)
for r in results:
r["retrieval_method"] = "cypher"
return results
except Exception:
return []
async def _generate_cypher(self, query: str) -> str:
schema_info = ""
if self.ontology:
schema_info = f"""
Graph Schema:
- Entity Types: {', '.join(self.ontology.entity_types)}
- Relationship Types: {', '.join(self.ontology.relationship_types)}
"""
prompt = f"""
You are a Cypher query generator. Generate a Cypher query for Neo4j.
Question: {query}
{schema_info}
Rules:
1. Only use entity labels and relationship types from the schema
2. Use MATCH clauses to find patterns
3. Use WHERE clauses for filtering
4. Return relevant data with RETURN clause
5. Add LIMIT 20 to prevent excessive results
6. Do not use deprecated syntax
Return only the Cypher query, no explanation.
"""
response = await self.llm.complete(
prompt,
system_prompt="You generate syntactically correct Cypher queries for Neo4j.",
temperature=0.1
)
cypher = response.strip()
if "```cypher" in cypher:
cypher = cypher.split("```cypher")[1].split("```")[0]
elif "```" in cypher:
cypher = cypher.split("```")[1].split("```")[0]
return cypher.strip()
def _validate_cypher(self, cypher: str) -> bool:
if not cypher:
return False
cypher_upper = cypher.upper()
if "MATCH" not in cypher_upper and "CALL" not in cypher_upper:
return False
dangerous_keywords = ["DELETE", "DETACH DELETE", "DROP", "REMOVE"]
for keyword in dangerous_keywords:
if keyword in cypher_upper:
return False
return True
async def _correct_cypher(self, cypher: str, query: str) -> str:
prompt = f"""
This Cypher query may have issues:
{cypher}
Original question: {query}
Fix any syntax errors or schema violations. Return only the corrected Cypher query.
"""
response = await self.llm.complete(prompt, temperature=0.1)
corrected = response.strip()
if "```" in corrected:
corrected = corrected.split("```")[1]
if corrected.startswith("cypher"):
corrected = corrected[6:]
corrected = corrected.split("```")[0]
return corrected.strip()
async def _correct_cypher_with_error(self, cypher: str, query: str, error: str) -> str:
prompt = f"""
This Cypher query failed with an error:
Query: {cypher}
Error: {error}
Original question: {query}
Fix the query to resolve the error. Return only the corrected Cypher.
"""
response = await self.llm.complete(prompt, temperature=0.1)
corrected = response.strip()
if "```" in corrected:
corrected = corrected.split("```")[1]
if corrected.startswith("cypher"):
corrected = corrected[6:]
corrected = corrected.split("```")[0]
return corrected.strip()
# ββ Metadata Filter Tool βββββββββββββββββββββββββββββββββββββββββββββββββββββ
class MetadataFilterTool:
"""Filter-based retrieval using metadata constraints"""
def __init__(self, store: Neo4jStore):
self.store = store
self.name = "metadata_filter"
self.description = "Filter entities or chunks by metadata attributes (date, type, source, etc.)"
async def run(
self,
filters: Dict[str, Any],
limit: int = 20
) -> List[Dict[str, Any]]:
import re as _re
# SECURITY: allowlist property key characters to prevent Cypher injection.
# Keys must match Neo4j property naming rules: letters, digits, underscores only.
_SAFE_KEY = _re.compile(r'^[A-Za-z_][A-Za-z0-9_]*$')
where_clauses = []
params: Dict[str, Any] = {}
for i, (key, value) in enumerate(filters.items()):
if not _SAFE_KEY.match(key):
# Skip keys that don't match the allowlist β never interpolate them
continue
param_name = f"param_{i}"
if isinstance(value, list):
where_clauses.append(f"n.{key} IN ${param_name}")
else:
where_clauses.append(f"n.{key} = ${param_name}")
params[param_name] = value
where_clause = " AND ".join(where_clauses) if where_clauses else "true"
query = f"""
MATCH (n)
WHERE {where_clause}
RETURN n
LIMIT $limit
"""
params["limit"] = limit
results = await self.store.execute_query(query, params)
for r in results:
r["retrieval_method"] = "metadata_filter"
return results
# ββ Gap #4: LLM-as-a-Judge Scorer ββββββββββββββββββββββββββββββββββββββββββββ
class LLMJudge:
"""
LLM-as-a-Judge for real confidence scoring.
Replaces the fake len(contexts)/top_k formula.
"""
def __init__(self, llm: UnifiedLLMProvider):
self.llm = llm
async def score(
self,
query: str,
answer: str,
contexts: List[Dict[str, Any]]
) -> Dict[str, Any]:
"""
Evaluate how well the answer is grounded in the retrieved contexts.
Returns:
{
"score": float (0.0-1.0),
"reasoning": str,
"grounded_claims": int,
"ungrounded_claims": int,
"hallucination_risk": "low" | "medium" | "high"
}
"""
if not contexts or not answer:
return {
"score": 0.0,
"reasoning": "No contexts or empty answer",
"grounded_claims": 0,
"ungrounded_claims": 0,
"hallucination_risk": "high"
}
context_text = "\n\n".join([
f"[Context {i+1}]: {ctx.get('text', str(ctx))[:400]}"
for i, ctx in enumerate(contexts[:6])
])
prompt = f"""You are an expert fact-checker evaluating whether an AI answer is grounded in source documents.
Question: {query}
AI Answer:
{answer[:600]}
Source Contexts:
{context_text}
Evaluate the answer. For each claim in the answer, check if it is directly supported by at least one context.
Rate the answer:
- 0.9-1.0: Every claim is clearly stated in the contexts (no hallucination)
- 0.7-0.9: Most claims supported, minor paraphrasing only
- 0.5-0.7: Some claims supported, some come from model training/inference
- 0.3-0.5: Many claims not traceable to contexts
- 0.0-0.3: Answer largely contradicts or ignores provided contexts
Return ONLY valid JSON (no markdown, no extra text):
{{"score": 0.85, "reasoning": "brief explanation", "grounded_claims": 4, "ungrounded_claims": 1}}"""
try:
response = await self.llm.complete(
prompt,
system_prompt="You are a strict factuality evaluator. Return only valid JSON.",
temperature=settings.judge_temperature
)
cleaned = response.strip()
# Strip any markdown code fences
if "```" in cleaned:
cleaned = cleaned.split("```")[1]
if cleaned.startswith("json"):
cleaned = cleaned[4:]
cleaned = cleaned.split("```")[0]
data = json.loads(cleaned.strip())
score = float(data.get("score", 0.5))
grounded = int(data.get("grounded_claims", 0))
ungrounded = int(data.get("ungrounded_claims", 0))
if score >= 0.75:
risk = "low"
elif score >= 0.5:
risk = "medium"
else:
risk = "high"
return {
"score": score,
"reasoning": data.get("reasoning", ""),
"grounded_claims": grounded,
"ungrounded_claims": ungrounded,
"hallucination_risk": risk
}
except Exception as e:
logger.info(f"LLM Judge error: {e}")
# Fallback: use a simple heuristic
base_score = min(1.0, len(contexts) / max(settings.default_top_k, 1))
return {
"score": base_score,
"reasoning": f"Heuristic score (judge unavailable: {e})",
"grounded_claims": len(contexts),
"ungrounded_claims": 0,
"hallucination_risk": "medium" if base_score < 0.7 else "low"
}
# ββ Gap #8: RAG Evaluator ββββββββββββββββββββββββββββββββββββββββββββββββββββ
class RAGEvaluator:
"""
RAGAS-style evaluation metrics for the quality dashboard.
Computes faithfulness, answer_relevancy, context_precision.
"""
def __init__(self, llm: UnifiedLLMProvider):
self.llm = llm
async def evaluate(
self,
question: str,
answer: str,
contexts: List[str],
ground_truth: Optional[str] = None
) -> Dict[str, float]:
"""
Run all evaluation metrics in parallel.
Returns dict with metric scores.
"""
tasks = [
self._faithfulness(answer, contexts),
self._answer_relevancy(question, answer),
self._context_precision(question, contexts),
]
results = await asyncio.gather(*tasks, return_exceptions=True)
faithfulness = results[0] if not isinstance(results[0], Exception) else 0.5
relevancy = results[1] if not isinstance(results[1], Exception) else 0.5
precision = results[2] if not isinstance(results[2], Exception) else 0.5
overall = (faithfulness * 0.4 + relevancy * 0.35 + precision * 0.25)
return {
"faithfulness": faithfulness,
"answer_relevancy": relevancy,
"context_precision": precision,
"overall_score": overall,
"hallucination_detected": faithfulness < 0.5
}
async def _faithfulness(self, answer: str, contexts: List[str]) -> float:
"""
Measure: Are all claims in the answer supported by the contexts?
"""
ctx_text = "\n".join([f"- {c[:300]}" for c in contexts[:5]])
prompt = f"""Given these source contexts:
{ctx_text}
And this answer:
{answer[:400]}
Score from 0.0 to 1.0 how faithfully the answer is grounded in the contexts.
1.0 = every statement is supported. 0.0 = answer contradicts or ignores contexts.
Return ONLY a float number like: 0.82"""
try:
resp = await self.llm.complete(prompt, temperature=0.0)
return float(resp.strip().split()[0])
except Exception:
return 0.5
async def _answer_relevancy(self, question: str, answer: str) -> float:
"""
Measure: Does the answer actually address the question?
"""
prompt = f"""Question: {question}
Answer: {answer[:400]}
Score from 0.0 to 1.0 how relevant and complete the answer is for the question.
1.0 = fully answers the question. 0.0 = completely off-topic.
Return ONLY a float number like: 0.91"""
try:
resp = await self.llm.complete(prompt, temperature=0.0)
return float(resp.strip().split()[0])
except Exception:
return 0.5
async def _context_precision(self, question: str, contexts: List[str]) -> float:
"""
Measure: Are the retrieved contexts useful for answering the question?
"""
if not contexts:
return 0.0
ctx_text = "\n".join([f"[{i+1}]: {c[:200]}" for i, c in enumerate(contexts[:5])])
prompt = f"""Question: {question}
Retrieved contexts:
{ctx_text}
Score from 0.0 to 1.0 how relevant and useful these contexts are for answering the question.
1.0 = all contexts are highly relevant. 0.0 = all contexts are irrelevant.
Return ONLY a float number like: 0.75"""
try:
resp = await self.llm.complete(prompt, temperature=0.0)
return float(resp.strip().split()[0])
except Exception:
return 0.5
# ββ MiroFish EntityEnricher: Entity Summary Search Tool βββββββββββββββββββββ
class EntitySummarySearchTool:
"""
Searches entity-level LLM summaries (from EntityEnricher) as a second
retrieval layer alongside chunk-level vector/BM25 search.
When entity summaries exist (e.summary IS NOT NULL), this tool embeds
the query and searches against them via BM25 text fallback or filters.
Complements HybridSearchTool which searches raw chunk text.
"""
def __init__(self, store: Neo4jStore, llm: UnifiedLLMProvider) -> None:
self.store = store
self.llm = llm
self.name = "entity_summary_search"
self.description = (
"Searches enriched entity profile summaries. "
"Best for 'what is X?', 'tell me about Y', named-entity questions."
)
async def run(
self,
query: str,
k: int = None,
entity_type: Optional[str] = None,
) -> List[Dict[str, Any]]:
"""
Find entities whose summaries are relevant to the query.
Uses fulltext search on the chunk index to surface entity names,
then fetches their summaries. Falls back to a simple CONTAINS match
if fulltext gives no results.
Args:
query: Natural language query
k: Max results (default from settings)
entity_type: Optional filter by entity type
Returns:
List of context dicts with entity name, type, summary text
"""
k = k or settings.default_top_k
type_filter = "AND e.type = $entity_type" if entity_type else ""
params: Dict[str, Any] = {"limit": k}
if entity_type:
params["entity_type"] = entity_type
# Strategy 1: Use BM25 index to find relevant entities via chunkβentity links
results: List[Dict[str, Any]] = []
try:
bm25_cypher = f"""
CALL db.index.fulltext.queryNodes('chunk_text_index', $query)
YIELD node, score
MATCH (node)-[:MENTIONS]->(e:Entity)
WHERE e.summary IS NOT NULL AND e.summary <> ''
{type_filter}
RETURN DISTINCT e.name as name, e.type as type, e.summary as summary,
score
ORDER BY score DESC
LIMIT $limit
"""
params["query"] = query
rows = await self.store.execute_query(bm25_cypher, params)
for r in rows:
results.append({
"id": f"entity_summary:{r['name']}",
"text": f"[Entity: {r['name']} ({r['type']})] {r['summary']}",
"entity_name": r["name"],
"entity_type": r.get("type", "Entity"),
"summary": r["summary"],
"score": r.get("score", 0.8),
"retrieval_method": "entity_summary",
})
except Exception as exc:
logger.info(f"[EntitySummarySearch BM25] {exc}")
# Strategy 2: Fallback β keyword match on entity names in graph
if not results:
try:
import re as _re
# SECURITY: sanitize each word before string-interpolating into Cypher.
# Allow basic punctuation to support terms like C++, Node.js, O'Connor.
_safe_word = _re.compile(r"^[\w\s\-\.\+\'\"]+$")
words = [
w for w in query.split()
if len(w) > 3 and _safe_word.match(w)
][:5]
if words:
# Use parameterised toLower comparison β one param per word
conditions_parts = []
for j, w in enumerate(words):
p = f"word_{j}"
conditions_parts.append(f"toLower(e.name) CONTAINS ${p}")
params[p] = w.lower()
conditions = " OR ".join(conditions_parts)
fallback_cypher = f"""
MATCH (e:Entity)
WHERE ({conditions}) AND e.summary IS NOT NULL AND e.summary <> ''
{type_filter}
RETURN e.name as name, e.type as type, e.summary as summary
LIMIT $limit
"""
rows = await self.store.execute_query(fallback_cypher, params)
for r in rows:
results.append({
"id": f"entity_summary:{r['name']}",
"text": f"[Entity: {r['name']} ({r['type']})] {r['summary']}",
"entity_name": r["name"],
"entity_type": r.get("type", "Entity"),
"summary": r["summary"],
"score": 0.6,
"retrieval_method": "entity_summary_fallback",
})
except Exception as exc:
logger.info(f"[EntitySummarySearch fallback] {exc}")
return results[:k]
|