fleet-memory/hindsight-api/hindsight_api/engine/search/retrieval.py
Nicolò Boschi 3bb0a58ded
Increase graph neighbor limit and benchmark improvements (#18)
* Improve LongMemEval benchmark with structured prompts and better options

- Add --context-format option with 'json' (original) and 'structured' modes
- Structured format groups facts with source chunks for better LLM comprehension
- Add detailed instructions for date calculations, relative time handling, and abstention
- Add --source-results flag to read failed questions from a different file
- Allow --category to be combined with --max-instances for sampling
- Fix Gemini structured output by passing response_schema parameter
- Add retry logic for empty Gemini responses with block reason logging
- Add judge prompt comparison documentation

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* fix recall in benchmarks

* Improve LongMemEval prompt and Gemini error handling

- Add JSONDecodeError retry for Gemini truncated responses
- Increase max_tokens to 32768 for thinking models
- Add counting/disambiguation guidance to structured prompt
- Add "when in doubt, undercount" and overlap detection rules

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* Add connection error retry and preference question guidance

- Add APIConnectionError retry for OpenAI client (server disconnects)
- Add recommendation/preference question guidance to structured prompt
- Instruct model to build on user's existing tools/experiences

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* Make reasoning optional

* Seed for LLM through Groq

* fix entity and observations

* Increase graph retrieval neighbor limit for expanded entities

Doubled the neighbor limit multiplier from 10 to 20 in graph retrieval.
With expanded entity extraction (now including objects and concepts like
"kitchen"), facts share more common entities, causing the previous limit
to arbitrarily exclude relevant results. This fix ensures better recall
for questions about related items (e.g., kitchen items).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

* Expand entity extraction to include objects and concepts

Updated entity extraction prompt to include:
- Specific objects (coffee maker, toaster, car, laptop, kitchen)
- Abstract concepts/themes (friendship, career growth, loss, celebration)
- Places and organizations (IKEA, Goodwill, New York)

This enables better fact linking through shared entities. For example,
kitchen appliances now share a "kitchen" entity, allowing graph traversal
to find related facts like "replaced coffee maker" when querying about
"kitchen items".

Works in conjunction with the increased neighbor limit to improve recall.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Chris Bartholomew <chris.bartholomew@vectorize.io>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: andrew <andrew.neeser@me.com>
2025-12-08 15:24:13 +01:00

503 lines
19 KiB
Python

"""
Retrieval module for 4-way parallel search.
Implements:
1. Semantic retrieval (vector similarity)
2. BM25 retrieval (keyword/full-text search)
3. Graph retrieval (spreading activation)
4. Temporal retrieval (time-aware search with spreading)
"""
from typing import List, Dict, Any, Tuple, Optional
from datetime import datetime
import asyncio
from ..db_utils import acquire_with_retry
from .types import RetrievalResult
async def retrieve_semantic(
conn,
query_emb_str: str,
bank_id: str,
fact_type: str,
limit: int
) -> List[RetrievalResult]:
"""
Semantic retrieval via vector similarity.
Args:
conn: Database connection
query_emb_str: Query embedding as string
agent_id: bank ID
fact_type: Fact type to filter
limit: Maximum results to return
Returns:
List of RetrievalResult objects
"""
results = await conn.fetch(
"""
SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
1 - (embedding <=> $1::vector) AS similarity
FROM memory_units
WHERE bank_id = $2
AND embedding IS NOT NULL
AND fact_type = $3
AND (1 - (embedding <=> $1::vector)) >= 0.3
ORDER BY embedding <=> $1::vector
LIMIT $4
""",
query_emb_str, bank_id, fact_type, limit
)
return [RetrievalResult.from_db_row(dict(r)) for r in results]
async def retrieve_bm25(
conn,
query_text: str,
bank_id: str,
fact_type: str,
limit: int
) -> List[RetrievalResult]:
"""
BM25 keyword retrieval via full-text search.
Args:
conn: Database connection
query_text: Query text
agent_id: bank ID
fact_type: Fact type to filter
limit: Maximum results to return
Returns:
List of RetrievalResult objects
"""
import re
# Sanitize query text: remove special characters that have meaning in tsquery
# Keep only alphanumeric characters and spaces
sanitized_text = re.sub(r'[^\w\s]', ' ', query_text.lower())
# Split and filter empty strings
tokens = [token for token in sanitized_text.split() if token]
if not tokens:
# If no valid tokens, return empty results
return []
# Convert query to tsquery using OR for more flexible matching
# This prevents empty results when some terms are missing
query_tsquery = " | ".join(tokens)
results = await conn.fetch(
"""
SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
ts_rank_cd(search_vector, to_tsquery('english', $1)) AS bm25_score
FROM memory_units
WHERE bank_id = $2
AND fact_type = $3
AND search_vector @@ to_tsquery('english', $1)
ORDER BY bm25_score DESC
LIMIT $4
""",
query_tsquery, bank_id, fact_type, limit
)
return [RetrievalResult.from_db_row(dict(r)) for r in results]
async def retrieve_graph(
conn,
query_emb_str: str,
bank_id: str,
fact_type: str,
budget: int
) -> List[RetrievalResult]:
"""
Graph retrieval via spreading activation.
Args:
conn: Database connection
query_emb_str: Query embedding as string
agent_id: bank ID
fact_type: Fact type to filter
budget: Node budget for graph traversal
Returns:
List of RetrievalResult objects
"""
# Find entry points
entry_points = await conn.fetch(
"""
SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
1 - (embedding <=> $1::vector) AS similarity
FROM memory_units
WHERE bank_id = $2
AND embedding IS NOT NULL
AND fact_type = $3
AND (1 - (embedding <=> $1::vector)) >= 0.5
ORDER BY embedding <=> $1::vector
LIMIT 5
""",
query_emb_str, bank_id, fact_type
)
if not entry_points:
return []
# BFS-style spreading activation with batched neighbor fetching
visited = set()
results = []
queue = [(RetrievalResult.from_db_row(dict(r)), r["similarity"]) for r in entry_points]
budget_remaining = budget
# Process nodes in batches to reduce DB roundtrips
batch_size = 20 # Fetch neighbors for up to 20 nodes at once
while queue and budget_remaining > 0:
# Collect a batch of nodes to process
batch_nodes = []
batch_activations = {}
while queue and len(batch_nodes) < batch_size and budget_remaining > 0:
current, activation = queue.pop(0)
unit_id = current.id
if unit_id not in visited:
visited.add(unit_id)
budget_remaining -= 1
results.append(current)
batch_nodes.append(current.id)
batch_activations[unit_id] = activation
# Batch fetch neighbors for all nodes in this batch
# Fetch top weighted neighbors (batch_size * 20 = ~400 for good distribution)
if batch_nodes and budget_remaining > 0:
max_neighbors = len(batch_nodes) * 20
neighbors = await conn.fetch(
"""
SELECT mu.id, mu.text, mu.context, mu.occurred_start, mu.occurred_end, mu.mentioned_at,
mu.access_count, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id,
ml.weight, ml.link_type, ml.from_unit_id
FROM memory_links ml
JOIN memory_units mu ON ml.to_unit_id = mu.id
WHERE ml.from_unit_id = ANY($1::uuid[])
AND ml.weight >= 0.1
AND mu.fact_type = $2
ORDER BY ml.weight DESC
LIMIT $3
""",
batch_nodes, fact_type, max_neighbors
)
for n in neighbors:
neighbor_id = str(n["id"])
if neighbor_id not in visited:
# Get parent activation
parent_id = str(n["from_unit_id"])
activation = batch_activations.get(parent_id, 0.5)
# Boost activation for causal links (they're high-value relationships)
link_type = n["link_type"]
base_weight = n["weight"]
# Causal links get 1.5-2.0x boost depending on type
if link_type in ("causes", "caused_by"):
# Direct causation - very strong relationship
causal_boost = 2.0
elif link_type in ("enables", "prevents"):
# Conditional causation - strong but not as direct
causal_boost = 1.5
else:
# Temporal, semantic, entity links - standard weight
causal_boost = 1.0
effective_weight = base_weight * causal_boost
new_activation = activation * effective_weight * 0.8
if new_activation > 0.1:
neighbor_result = RetrievalResult.from_db_row(dict(n))
queue.append((neighbor_result, new_activation))
return results
async def retrieve_temporal(
conn,
query_emb_str: str,
bank_id: str,
fact_type: str,
start_date: datetime,
end_date: datetime,
budget: int,
semantic_threshold: float = 0.1
) -> List[RetrievalResult]:
"""
Temporal retrieval with spreading activation.
Strategy:
1. Find entry points (facts in date range with semantic relevance)
2. Spread through temporal links to related facts
3. Score by temporal proximity + semantic similarity + link weight
Args:
conn: Database connection
query_emb_str: Query embedding as string
agent_id: bank ID
fact_type: Fact type to filter
start_date: Start of time range
end_date: End of time range
budget: Node budget for spreading
semantic_threshold: Minimum semantic similarity to include
Returns:
List of RetrievalResult objects with temporal scores
"""
from datetime import timezone
# Ensure start_date and end_date are timezone-aware (UTC) to match database datetimes
if start_date.tzinfo is None:
start_date = start_date.replace(tzinfo=timezone.utc)
if end_date.tzinfo is None:
end_date = end_date.replace(tzinfo=timezone.utc)
entry_points = await conn.fetch(
"""
SELECT id, text, context, event_date, occurred_start, occurred_end, mentioned_at, access_count, embedding, fact_type, document_id, chunk_id,
1 - (embedding <=> $1::vector) AS similarity
FROM memory_units
WHERE bank_id = $2
AND fact_type = $3
AND embedding IS NOT NULL
AND (
-- Match if occurred range overlaps with query range
(occurred_start IS NOT NULL AND occurred_end IS NOT NULL
AND occurred_start <= $5 AND occurred_end >= $4)
OR
-- Match if mentioned_at falls within query range
(mentioned_at IS NOT NULL AND mentioned_at BETWEEN $4 AND $5)
OR
-- Match if any occurred date is set and overlaps (even if only start or end is set)
(occurred_start IS NOT NULL AND occurred_start BETWEEN $4 AND $5)
OR
(occurred_end IS NOT NULL AND occurred_end BETWEEN $4 AND $5)
)
AND (1 - (embedding <=> $1::vector)) >= $6
ORDER BY COALESCE(occurred_start, mentioned_at, occurred_end) DESC, (embedding <=> $1::vector) ASC
LIMIT 10
""",
query_emb_str, bank_id, fact_type, start_date, end_date, semantic_threshold
)
if not entry_points:
return []
# Calculate temporal scores for entry points
total_days = (end_date - start_date).total_seconds() / 86400
mid_date = start_date + (end_date - start_date) / 2 # Calculate once for all comparisons
results = []
visited = set()
for ep in entry_points:
unit_id = str(ep["id"])
visited.add(unit_id)
# Calculate temporal proximity using the most relevant date
# Priority: occurred_start/end (event time) > mentioned_at (mention time)
best_date = None
if ep["occurred_start"] is not None and ep["occurred_end"] is not None:
# Use midpoint of occurred range
best_date = ep["occurred_start"] + (ep["occurred_end"] - ep["occurred_start"]) / 2
elif ep["occurred_start"] is not None:
best_date = ep["occurred_start"]
elif ep["occurred_end"] is not None:
best_date = ep["occurred_end"]
elif ep["mentioned_at"] is not None:
best_date = ep["mentioned_at"]
# Temporal proximity score (closer to range center = higher score)
if best_date:
days_from_mid = abs((best_date - mid_date).total_seconds() / 86400)
temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
else:
temporal_proximity = 0.5 # Fallback if no dates (shouldn't happen due to WHERE clause)
# Create RetrievalResult with temporal scores
ep_result = RetrievalResult.from_db_row(dict(ep))
ep_result.temporal_score = temporal_proximity
ep_result.temporal_proximity = temporal_proximity
results.append(ep_result)
# Spread through temporal links
queue = [(RetrievalResult.from_db_row(dict(ep)), ep["similarity"], 1.0) for ep in entry_points] # (unit, semantic_sim, temporal_score)
budget_remaining = budget - len(entry_points)
while queue and budget_remaining > 0:
current, semantic_sim, temporal_score = queue.pop(0)
current_id = current.id
# Get neighbors via temporal and causal links
if budget_remaining > 0:
neighbors = await conn.fetch(
"""
SELECT mu.id, mu.text, mu.context, mu.event_date, mu.occurred_start, mu.occurred_end, mu.mentioned_at, mu.access_count, mu.embedding, mu.fact_type, mu.document_id, mu.chunk_id,
ml.weight, ml.link_type,
1 - (mu.embedding <=> $1::vector) AS similarity
FROM memory_links ml
JOIN memory_units mu ON ml.to_unit_id = mu.id
WHERE ml.from_unit_id = $2
AND ml.link_type IN ('temporal', 'causes', 'caused_by', 'enables', 'prevents')
AND ml.weight >= 0.1
AND mu.fact_type = $3
AND mu.embedding IS NOT NULL
AND (1 - (mu.embedding <=> $1::vector)) >= $4
ORDER BY ml.weight DESC
LIMIT 10
""",
query_emb_str, current.id, fact_type, semantic_threshold
)
for n in neighbors:
neighbor_id = str(n["id"])
if neighbor_id in visited:
continue
visited.add(neighbor_id)
budget_remaining -= 1
# Calculate temporal score for neighbor using best available date
neighbor_best_date = None
if n["occurred_start"] is not None and n["occurred_end"] is not None:
neighbor_best_date = n["occurred_start"] + (n["occurred_end"] - n["occurred_start"]) / 2
elif n["occurred_start"] is not None:
neighbor_best_date = n["occurred_start"]
elif n["occurred_end"] is not None:
neighbor_best_date = n["occurred_end"]
elif n["mentioned_at"] is not None:
neighbor_best_date = n["mentioned_at"]
if neighbor_best_date:
days_from_mid = abs((neighbor_best_date - mid_date).total_seconds() / 86400)
neighbor_temporal_proximity = 1.0 - min(days_from_mid / (total_days / 2), 1.0) if total_days > 0 else 1.0
else:
neighbor_temporal_proximity = 0.3 # Lower score if no temporal data
# Boost causal links (same as graph retrieval)
link_type = n["link_type"]
if link_type in ("causes", "caused_by"):
causal_boost = 2.0
elif link_type in ("enables", "prevents"):
causal_boost = 1.5
else:
causal_boost = 1.0
# Propagate temporal score through links (decay, with causal boost)
propagated_temporal = temporal_score * n["weight"] * causal_boost * 0.7
# Combined temporal score
combined_temporal = max(neighbor_temporal_proximity, propagated_temporal)
# Create RetrievalResult with temporal scores
neighbor_result = RetrievalResult.from_db_row(dict(n))
neighbor_result.temporal_score = combined_temporal
neighbor_result.temporal_proximity = neighbor_temporal_proximity
results.append(neighbor_result)
# Add to queue for further spreading
if budget_remaining > 0 and combined_temporal > 0.2:
queue.append((neighbor_result, n["similarity"], combined_temporal))
if budget_remaining <= 0:
break
return results
async def retrieve_parallel(
pool,
query_text: str,
query_embedding_str: str,
bank_id: str,
fact_type: str,
thinking_budget: int,
question_date: Optional[datetime] = None,
query_analyzer: Optional["QueryAnalyzer"] = None
) -> Tuple[List[RetrievalResult], List[RetrievalResult], List[RetrievalResult], Optional[List[RetrievalResult]], Dict[str, float], Optional[Tuple[datetime, datetime]]]:
"""
Run 3-way or 4-way parallel retrieval (adds temporal if detected).
Args:
pool: Database connection pool
query_text: Query text
query_embedding_str: Query embedding as string
agent_id: bank ID
fact_type: Fact type to filter
thinking_budget: Budget for graph traversal and retrieval limits
question_date: Optional date when question was asked (for temporal filtering)
query_analyzer: Query analyzer to use (defaults to TransformerQueryAnalyzer)
Returns:
Tuple of (semantic_results, bm25_results, graph_results, temporal_results, timings, temporal_constraint)
Each results list contains RetrievalResult objects
temporal_results is None if no temporal constraint detected
timings is a dict with per-method latencies in seconds
temporal_constraint is the (start_date, end_date) tuple if detected, else None
"""
# Detect temporal constraint
from .temporal_extraction import extract_temporal_constraint
import time
temporal_constraint = extract_temporal_constraint(
query_text, reference_date=question_date, analyzer=query_analyzer
)
# Wrapper to track timing for each retrieval method
async def timed_retrieval(name: str, coro):
start = time.time()
result = await coro
duration = time.time() - start
return result, name, duration
async def run_semantic():
async with acquire_with_retry(pool) as conn:
return await retrieve_semantic(conn, query_embedding_str, bank_id, fact_type, limit=thinking_budget)
async def run_bm25():
async with acquire_with_retry(pool) as conn:
return await retrieve_bm25(conn, query_text, bank_id, fact_type, limit=thinking_budget)
async def run_graph():
async with acquire_with_retry(pool) as conn:
return await retrieve_graph(conn, query_embedding_str, bank_id, fact_type, budget=thinking_budget)
async def run_temporal(start_date, end_date):
async with acquire_with_retry(pool) as conn:
return await retrieve_temporal(
conn, query_embedding_str, bank_id, fact_type,
start_date, end_date, budget=thinking_budget, semantic_threshold=0.1
)
# Run retrievals in parallel with timing
timings = {}
if temporal_constraint:
start_date, end_date = temporal_constraint
results = await asyncio.gather(
timed_retrieval("semantic", run_semantic()),
timed_retrieval("bm25", run_bm25()),
timed_retrieval("graph", run_graph()),
timed_retrieval("temporal", run_temporal(start_date, end_date))
)
semantic_results, _, timings["semantic"] = results[0]
bm25_results, _, timings["bm25"] = results[1]
graph_results, _, timings["graph"] = results[2]
temporal_results, _, timings["temporal"] = results[3]
else:
results = await asyncio.gather(
timed_retrieval("semantic", run_semantic()),
timed_retrieval("bm25", run_bm25()),
timed_retrieval("graph", run_graph())
)
semantic_results, _, timings["semantic"] = results[0]
bm25_results, _, timings["bm25"] = results[1]
graph_results, _, timings["graph"] = results[2]
temporal_results = None
return semantic_results, bm25_results, graph_results, temporal_results, timings, temporal_constraint