adai/MemoryBear
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(memory): add async user metadata extraction pipeline

Browse Source 
- Add MetadataExtractor to collect user-related statements post-dedup
  and extract profile/behavioral metadata via independent LLM call
- Add Celery task (extract_user_metadata) routed to memory_tasks queue
- Add metadata models (UserMetadata, UserMetadataProfile, etc.)
- Add metadata utility functions (clean, validate, merge with _op support)
- Add Jinja2 prompt template for metadata extraction (zh/en)
- Fix Lucene query parameter naming: rename `q` to `query` across all
  Cypher queries, graph_search functions, and callers
- Escape `/` in Lucene queries to prevent TokenMgrError
- Add `speaker` field to ChunkNode and persist it in Neo4j
- Remove unused imports (argparse, os, UUID) in search.py
- Fix unnecessary db context nesting in interest distribution task
This commit is contained in:
lanceyq
2026-04-09 11:01:56 +08:00
parent cfbf83f71e
commit f2d7479229
17 changed files with 714 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
api/app/repositories/neo4j/cypher_queries.py
View File

@@ -23,6 +23,7 @@ SET s += {
end_user_id: statement.end_user_id,
stmt_type: statement.stmt_type,
statement: statement.statement,
speaker: statement.speaker,
emotion_intensity: statement.emotion_intensity,
emotion_target: statement.emotion_target,
emotion_subject: statement.emotion_subject,
@@ -56,6 +57,7 @@ SET c += {
expired_at: chunk.expired_at,
dialog_id: chunk.dialog_id,
content: chunk.content,
speaker: chunk.speaker,
chunk_embedding: chunk.chunk_embedding,
sequence_number: chunk.sequence_number,
start_index: chunk.start_index,
@@ -283,7 +285,7 @@ LIMIT $limit
"""
SEARCH_STATEMENTS_BY_KEYWORD = """
CALL db.index.fulltext.queryNodes("statementsFulltext", $q) YIELD node AS s, score
CALL db.index.fulltext.queryNodes("statementsFulltext", $query) YIELD node AS s, score
WHERE ($end_user_id IS NULL OR s.end_user_id = $end_user_id)
OPTIONAL MATCH (c:Chunk)-[:CONTAINS]->(s)
OPTIONAL MATCH (s)-[:REFERENCES_ENTITY]->(e:ExtractedEntity)
@@ -307,7 +309,7 @@ LIMIT $limit
"""
# 查询实体名称包含指定字符串的实体
SEARCH_ENTITIES_BY_NAME = """
CALL db.index.fulltext.queryNodes("entitiesFulltext", $q) YIELD node AS e, score
CALL db.index.fulltext.queryNodes("entitiesFulltext", $query) YIELD node AS e, score
WHERE ($end_user_id IS NULL OR e.end_user_id = $end_user_id)
OPTIONAL MATCH (s:Statement)-[:REFERENCES_ENTITY]->(e)
OPTIONAL MATCH (c:Chunk)-[:CONTAINS]->(s)
@@ -337,21 +339,21 @@ LIMIT $limit
"""
SEARCH_ENTITIES_BY_NAME_OR_ALIAS = """
CALL db.index.fulltext.queryNodes("entitiesFulltext", $q) YIELD node AS e, score
CALL db.index.fulltext.queryNodes("entitiesFulltext", $query) YIELD node AS e, score
WHERE ($end_user_id IS NULL OR e.end_user_id = $end_user_id)
WITH e, score
WITH collect({entity: e, score: score}) AS fulltextResults
With collect({entity: e, score: score}) AS fulltextResults
OPTIONAL MATCH (ae:ExtractedEntity)
WHERE ($end_user_id IS NULL OR ae.end_user_id = $end_user_id)
AND ae.aliases IS NOT NULL
AND ANY(alias IN ae.aliases WHERE toLower(alias) CONTAINS toLower($q))
AND ANY(alias IN ae.aliases WHERE toLower(alias) CONTAINS toLower($query))
WITH fulltextResults, collect(ae) AS aliasEntities
UNWIND (fulltextResults + [x IN aliasEntities | {entity: x, score:
CASE
WHEN ANY(alias IN x.aliases WHERE toLower(alias) = toLower($q)) THEN 1.0
WHEN ANY(alias IN x.aliases WHERE toLower(alias) STARTS WITH toLower($q)) THEN 0.9
WHEN ANY(alias IN x.aliases WHERE toLower(alias) = toLower($query)) THEN 1.0
WHEN ANY(alias IN x.aliases WHERE toLower(alias) STARTS WITH toLower($query)) THEN 0.9
ELSE 0.8
END
}]) AS row
@@ -384,7 +386,7 @@ LIMIT $limit
SEARCH_CHUNKS_BY_CONTENT = """
CALL db.index.fulltext.queryNodes("chunksFulltext", $q) YIELD node AS c, score
CALL db.index.fulltext.queryNodes("chunksFulltext", $query) YIELD node AS c, score
WHERE ($end_user_id IS NULL OR c.end_user_id = $end_user_id)
OPTIONAL MATCH (c)-[:CONTAINS]->(s:Statement)
OPTIONAL MATCH (s)-[:REFERENCES_ENTITY]->(e:ExtractedEntity)
@@ -501,7 +503,7 @@ LIMIT $limit
"""
SEARCH_STATEMENTS_BY_KEYWORD_TEMPORAL = """
CALL db.index.fulltext.queryNodes("statementsFulltext", $q) YIELD node AS s, score
CALL db.index.fulltext.queryNodes("statementsFulltext", $query) YIELD node AS s, score
WHERE ($end_user_id IS NULL OR s.end_user_id = $end_user_id)
AND ((($start_date IS NULL OR (s.created_at IS NOT NULL AND datetime(s.created_at) >= datetime($start_date)))
AND ($end_date IS NULL OR (s.created_at IS NOT NULL AND datetime(s.created_at) <= datetime($end_date))))
@@ -677,7 +679,7 @@ SET n.invalid_at = $new_invalid_at
# MemorySummary keyword search using fulltext index
SEARCH_MEMORY_SUMMARIES_BY_KEYWORD = """
CALL db.index.fulltext.queryNodes("summariesFulltext", $q) YIELD node AS m, score
CALL db.index.fulltext.queryNodes("summariesFulltext", $query) YIELD node AS m, score
WHERE ($end_user_id IS NULL OR m.end_user_id = $end_user_id)
OPTIONAL MATCH (m)-[:DERIVED_FROM_STATEMENT]->(s:Statement)
RETURN m.id AS id,
@@ -1363,7 +1365,7 @@ RETURN c.community_id AS community_id
# Community keyword search: matches name or summary via fulltext index
SEARCH_COMMUNITIES_BY_KEYWORD = """
CALL db.index.fulltext.queryNodes("communitiesFulltext", $q) YIELD node AS c, score
CALL db.index.fulltext.queryNodes("communitiesFulltext", $query) YIELD node AS c, score
WHERE ($end_user_id IS NULL OR c.end_user_id = $end_user_id)
RETURN c.community_id AS id,
c.name AS name,
@@ -1451,7 +1453,7 @@ RETURN elementId(r) AS uuid
"""
SEARCH_PERCEPTUAL_BY_KEYWORD = """
CALL db.index.fulltext.queryNodes("perceptualFulltext", $q) YIELD node AS p, score
CALL db.index.fulltext.queryNodes("perceptualFulltext", $query) YIELD node AS p, score
WHERE p.end_user_id = $end_user_id
RETURN p.id AS id,
p.end_user_id AS end_user_id,

49
api/app/repositories/neo4j/graph_search.py
View File

@@ -2,6 +2,7 @@ import asyncio
import logging
from typing import Any, Dict, List, Optional
from app.core.memory.utils.data.text_utils import escape_lucene_query
from app.repositories.neo4j.cypher_queries import (
CHUNK_EMBEDDING_SEARCH,
COMMUNITY_EMBEDDING_SEARCH,
@@ -87,7 +88,7 @@ async def _update_activation_values_batch(
unique_node_ids.append(node_id)
if not unique_node_ids:
logger.warning(f"批量更新激活值：没有有效的节点ID")
logger.warning("批量更新激活值：没有有效的节点ID")
return nodes
# 记录去重信息（仅针对具有有效 ID 的节点）
@@ -223,7 +224,7 @@ async def _update_search_results_activation(
async def search_graph(
connector: Neo4jConnector,
q: str,
query: str,
end_user_id: Optional[str] = None,
limit: int = 50,
include: List[str] = None,
@@ -234,14 +235,14 @@ async def search_graph(
OPTIMIZED: Runs all queries in parallel using asyncio.gather()
INTEGRATED: Updates activation values for knowledge nodes before returning results
- Statements: matches s.statement CONTAINS q
- Entities: matches e.name CONTAINS q
- Chunks: matches s.content CONTAINS q (from Statement nodes)
- Summaries: matches ms.content CONTAINS q
- Statements: matches s.statement CONTAINS query
- Entities: matches e.name CONTAINS query
- Chunks: matches s.content CONTAINS query (from Statement nodes)
- Summaries: matches ms.content CONTAINS query
Args:
connector: Neo4j connector
q: Query text
query: Query text for full-text search
end_user_id: Optional group filter
limit: Max results per category
include: List of categories to search (default: all)
@@ -252,6 +253,9 @@ async def search_graph(
if include is None:
include = ["statements", "chunks", "entities", "summaries"]
# Escape Lucene special characters to prevent query parse errors
escaped_query = escape_lucene_query(query)
# Prepare tasks for parallel execution
tasks = []
task_keys = []
@@ -260,7 +264,7 @@ async def search_graph(
tasks.append(connector.execute_query(
SEARCH_STATEMENTS_BY_KEYWORD,
json_format=True,
q=q,
query=escaped_query,
end_user_id=end_user_id,
limit=limit,
))
@@ -270,7 +274,7 @@ async def search_graph(
tasks.append(connector.execute_query(
SEARCH_ENTITIES_BY_NAME_OR_ALIAS,
json_format=True,
q=q,
query=escaped_query,
end_user_id=end_user_id,
limit=limit,
))
@@ -280,7 +284,7 @@ async def search_graph(
tasks.append(connector.execute_query(
SEARCH_CHUNKS_BY_CONTENT,
json_format=True,
q=q,
query=escaped_query,
end_user_id=end_user_id,
limit=limit,
))
@@ -290,7 +294,7 @@ async def search_graph(
tasks.append(connector.execute_query(
SEARCH_MEMORY_SUMMARIES_BY_KEYWORD,
json_format=True,
q=q,
query=escaped_query,
end_user_id=end_user_id,
limit=limit,
))
@@ -300,7 +304,7 @@ async def search_graph(
tasks.append(connector.execute_query(
SEARCH_COMMUNITIES_BY_KEYWORD,
json_format=True,
q=q,
query=escaped_query,
end_user_id=end_user_id,
limit=limit,
))
@@ -482,7 +486,7 @@ async def search_graph_by_embedding(
update_time = time.time() - update_start
logger.info(f"[PERF] Activation value updates took: {update_time:.4f}s")
else:
logger.info(f"[PERF] Skipping activation updates (only summaries)")
logger.info("[PERF] Skipping activation updates (only summaries)")
return results
@@ -520,7 +524,7 @@ async def get_dedup_candidates_for_entities( # 适配新版查询：使用全
# 全文索引按名称检索（包含 CONTAINS 语义）
rows = await connector.execute_query(
SEARCH_ENTITIES_BY_NAME,
q=name,
query=escape_lucene_query(name),
end_user_id=end_user_id,
limit=100,
)
@@ -544,7 +548,7 @@ async def get_dedup_candidates_for_entities( # 适配新版查询：使用全
try:
rows = await connector.execute_query(
SEARCH_ENTITIES_BY_NAME,
q=name.lower(),
query=escape_lucene_query(name.lower()),
end_user_id=end_user_id,
limit=100,
)
@@ -593,11 +597,12 @@ async def search_graph_by_keyword_temporal(
- Returns up to 'limit' statements
"""
if not query_text:
logger.warning(f"query_text不能为空")
logger.warning("query_text不能为空")
return {"statements": []}
escaped_query = escape_lucene_query(query_text)
statements = await connector.execute_query(
SEARCH_STATEMENTS_BY_KEYWORD_TEMPORAL,
q=query_text,
query=escaped_query,
end_user_id=end_user_id,
start_date=start_date,
end_date=end_date,
@@ -671,7 +676,7 @@ async def search_graph_by_dialog_id(
- Returns up to 'limit' dialogues
"""
if not dialog_id:
logger.warning(f"dialog_id不能为空")
logger.warning("dialog_id不能为空")
return {"dialogues": []}
dialogues = await connector.execute_query(
@@ -690,7 +695,7 @@ async def search_graph_by_chunk_id(
limit: int = 1,
) -> Dict[str, List[Dict[str, Any]]]:
if not chunk_id:
logger.warning(f"chunk_id不能为空")
logger.warning("chunk_id不能为空")
return {"chunks": []}
chunks = await connector.execute_query(
SEARCH_CHUNK_BY_CHUNK_ID,
@@ -968,7 +973,7 @@ async def search_graph_l_valid_at(
async def search_perceptual(
connector: Neo4jConnector,
q: str,
query: str,
end_user_id: Optional[str] = None,
limit: int = 10,
) -> Dict[str, List[Dict[str, Any]]]:
@@ -979,7 +984,7 @@ async def search_perceptual(
Args:
connector: Neo4j connector
q: Query text
query: Query text for full-text search
end_user_id: Optional user filter
limit: Max results
@@ -989,7 +994,7 @@ async def search_perceptual(
try:
perceptuals = await connector.execute_query(
SEARCH_PERCEPTUAL_BY_KEYWORD,
q=q,
query=escape_lucene_query(query),
end_user_id=end_user_id,
limit=limit,
)