feat: auto-summarize trigger for agent memory
- Memory Service: POST /agents/{agent_id}/summarize endpoint
- Fetches recent events by agent_id (new db.list_facts_by_agent)
- Generates structured summary via DeepSeek LLM
- Saves summary to PostgreSQL facts + Qdrant vector store
- Returns structured JSON (summary, goals, decisions, key_facts)
- Gateway memory_client: auto-trigger after 30 turns
- Turn counter per chat (agent_id:channel_id)
- 5-minute debounce between summarize calls
- Fire-and-forget via asyncio.ensure_future (non-blocking)
- Configurable via SUMMARIZE_TURN_THRESHOLD / SUMMARIZE_DEBOUNCE_SECONDS
- Database: list_facts_by_agent() for agent-level queries without user_id
Tested on NODE1: Helion summarize returns valid Ukrainian summary with 20 events.
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Apple
@@ -517,6 +517,25 @@ class Database:
|
||||
|
||||
return [dict(row) for row in rows]
|
||||
|
||||
async def list_facts_by_agent(
|
||||
self,
|
||||
agent_id: str,
|
||||
channel_id: str = None,
|
||||
limit: int = 60
|
||||
) -> list:
|
||||
"""List facts for an agent (any user), ordered by most recent."""
|
||||
async with self.pool.acquire() as conn:
|
||||
query = "SELECT * FROM user_facts WHERE agent_id = $1"
|
||||
params = [agent_id]
|
||||
if channel_id:
|
||||
query += " AND fact_key LIKE '%' || $2 || '%'"
|
||||
params.append(channel_id)
|
||||
query += " ORDER BY updated_at DESC"
|
||||
query += f" LIMIT ${len(params) + 1}"
|
||||
params.append(limit)
|
||||
rows = await conn.fetch(query, *params)
|
||||
return [dict(row) for row in rows]
|
||||
|
||||
async def delete_fact(
|
||||
self,
|
||||
user_id: str,
|
||||
|
||||
@@ -1011,3 +1011,169 @@ async def get_stats():
|
||||
if __name__ == "__main__":
|
||||
import uvicorn
|
||||
uvicorn.run(app, host="0.0.0.0", port=8000)
|
||||
|
||||
|
||||
# ============================================================================
|
||||
# AGENT-LEVEL SUMMARIZE (called by Gateway auto-trigger)
|
||||
# ============================================================================
|
||||
|
||||
class AgentSummarizeRequest(BaseModel):
|
||||
"""Request to generate a summary of recent agent memory events"""
|
||||
channel_id: Optional[str] = None
|
||||
user_id: Optional[str] = None
|
||||
max_events: int = 60 # how many recent events to summarize
|
||||
force: bool = False # bypass debounce check
|
||||
|
||||
|
||||
@app.post("/agents/{agent_id}/summarize")
|
||||
async def summarize_agent_memory(agent_id: str, request: AgentSummarizeRequest):
|
||||
"""
|
||||
Generate rolling summary of recent agent memory events.
|
||||
Called by Gateway when conversation reaches threshold.
|
||||
|
||||
1. Fetch recent events from facts table (agent-isolated)
|
||||
2. Generate structured summary via DeepSeek LLM
|
||||
3. Save summary back as a special fact for future context retrieval
|
||||
4. Index summary in Qdrant for semantic search
|
||||
"""
|
||||
import json as json_lib
|
||||
from datetime import datetime as dt
|
||||
from uuid import uuid4
|
||||
|
||||
try:
|
||||
# 1. Fetch recent events for this agent (all users)
|
||||
facts = await db.list_facts_by_agent(
|
||||
agent_id=agent_id,
|
||||
channel_id=request.channel_id,
|
||||
limit=request.max_events
|
||||
)
|
||||
|
||||
# Filter for chat events (optionally by channel)
|
||||
events = []
|
||||
for fact in facts:
|
||||
if not fact.get("fact_key", "").startswith("chat_event:"):
|
||||
continue
|
||||
event_data = fact.get("fact_value_json", {})
|
||||
if isinstance(event_data, str):
|
||||
try:
|
||||
event_data = json_lib.loads(event_data)
|
||||
except Exception:
|
||||
event_data = {}
|
||||
if not isinstance(event_data, dict):
|
||||
event_data = {}
|
||||
if request.channel_id and event_data.get("channel_id") != request.channel_id:
|
||||
continue
|
||||
events.append(event_data)
|
||||
|
||||
if len(events) < 5:
|
||||
return {
|
||||
"status": "skipped",
|
||||
"reason": f"Too few events ({len(events)}), need at least 5",
|
||||
"events_count": len(events)
|
||||
}
|
||||
|
||||
# 2. Format events for LLM
|
||||
formatted_events = []
|
||||
for e in events:
|
||||
formatted_events.append({
|
||||
"role": e.get("role", "user"),
|
||||
"content": e.get("content", ""),
|
||||
"timestamp": e.get("timestamp", "")
|
||||
})
|
||||
|
||||
# 3. Generate summary via DeepSeek LLM
|
||||
llm_result = await _llm_generate_summary(formatted_events)
|
||||
|
||||
# 4. Save summary as a special fact
|
||||
summary_id = str(uuid4())
|
||||
timestamp = dt.utcnow().isoformat()
|
||||
summary_fact = {
|
||||
"type": "dialog_summary",
|
||||
"summary_id": summary_id,
|
||||
"agent_id": agent_id,
|
||||
"channel_id": request.channel_id,
|
||||
"summary": llm_result["summary"],
|
||||
"goals": llm_result["goals"],
|
||||
"decisions": llm_result["decisions"],
|
||||
"open_questions": llm_result["open_questions"],
|
||||
"next_steps": llm_result["next_steps"],
|
||||
"key_facts": llm_result["key_facts"],
|
||||
"events_summarized": len(events),
|
||||
"timestamp": timestamp
|
||||
}
|
||||
|
||||
await db.ensure_facts_table()
|
||||
await db.upsert_fact(
|
||||
user_id=request.user_id or "system",
|
||||
fact_key=f"summary:{agent_id}:{request.channel_id or all}:{timestamp}",
|
||||
fact_value_json=summary_fact,
|
||||
team_id=None,
|
||||
agent_id=agent_id
|
||||
)
|
||||
|
||||
# 5. Index in Qdrant for semantic search
|
||||
summary_text = llm_result["summary"]
|
||||
if summary_text and len(summary_text) > 20:
|
||||
try:
|
||||
from .embedding import get_document_embeddings
|
||||
from qdrant_client import models as qmodels
|
||||
|
||||
embeddings = await get_document_embeddings([summary_text])
|
||||
if embeddings:
|
||||
vector = embeddings[0]
|
||||
collection_name = f"{agent_id}_summaries"
|
||||
|
||||
# Ensure collection exists
|
||||
try:
|
||||
vector_store.client.get_collection(collection_name)
|
||||
except Exception:
|
||||
vector_store.client.create_collection(
|
||||
collection_name=collection_name,
|
||||
vectors_config=qmodels.VectorParams(
|
||||
size=len(vector),
|
||||
distance=qmodels.Distance.COSINE
|
||||
)
|
||||
)
|
||||
logger.info("created_summary_collection", collection=collection_name)
|
||||
|
||||
vector_store.client.upsert(
|
||||
collection_name=collection_name,
|
||||
points=[
|
||||
qmodels.PointStruct(
|
||||
id=summary_id,
|
||||
vector=vector,
|
||||
payload={
|
||||
"type": "dialog_summary",
|
||||
"agent_id": agent_id,
|
||||
"channel_id": request.channel_id,
|
||||
"events_count": len(events),
|
||||
"summary_text": summary_text,
|
||||
"timestamp": timestamp
|
||||
}
|
||||
)
|
||||
]
|
||||
)
|
||||
except Exception as ve:
|
||||
logger.warning("summary_qdrant_index_failed",
|
||||
error=str(ve), agent_id=agent_id)
|
||||
|
||||
logger.info("agent_summary_created",
|
||||
agent_id=agent_id,
|
||||
channel_id=request.channel_id,
|
||||
events_count=len(events),
|
||||
summary_len=len(summary_text))
|
||||
|
||||
return {
|
||||
"status": "ok",
|
||||
"summary_id": summary_id,
|
||||
"summary": llm_result["summary"],
|
||||
"goals": llm_result["goals"],
|
||||
"decisions": llm_result["decisions"],
|
||||
"key_facts": llm_result["key_facts"],
|
||||
"events_summarized": len(events),
|
||||
"timestamp": timestamp
|
||||
}
|
||||
|
||||
except Exception as e:
|
||||
logger.error("agent_summarize_failed", error=str(e), agent_id=agent_id)
|
||||
raise HTTPException(status_code=500, detail=str(e))
|
||||
|
||||
Reference in New Issue
Block a user