runtime: sync router/gateway/config policy and clan role registry
This commit is contained in:
Apple
@@ -363,6 +363,8 @@ class MemoryRetrieval:
|
||||
query: str,
|
||||
agent_id: str = "helion",
|
||||
platform_user_id: Optional[str] = None,
|
||||
chat_id: Optional[str] = None,
|
||||
user_id: Optional[str] = None,
|
||||
visibility: str = "platform",
|
||||
limit: int = 5
|
||||
) -> List[Dict[str, Any]]:
|
||||
@@ -377,6 +379,16 @@ class MemoryRetrieval:
|
||||
|
||||
all_results = []
|
||||
|
||||
q = (query or "").lower()
|
||||
# If user explicitly asks about documents/catalogs, prefer knowledge base docs over chat snippets.
|
||||
is_doc_query = any(k in q for k in ["pdf", "каталог", "каталоз", "документ", "файл", "стор", "page", "pages"])
|
||||
# Simple keyword gate to avoid irrelevant chat snippets dominating doc queries.
|
||||
# Example: when asking "з каталогу Defenda 2026 ... гліфосат", old "Бокаші" messages may match too well.
|
||||
topic_keywords: List[str] = []
|
||||
for kw in ["defenda", "ifagri", "bayer", "гліфосат", "glyphos", "глифос", "npk", "мінерал", "добрив", "гербіц", "фунгіц", "інсектиц"]:
|
||||
if kw in q:
|
||||
topic_keywords.append(kw)
|
||||
|
||||
# Dynamic collection names based on agent_id
|
||||
memory_items_collection = f"{agent_id}_memory_items"
|
||||
messages_collection = f"{agent_id}_messages"
|
||||
@@ -420,18 +432,34 @@ class MemoryRetrieval:
|
||||
|
||||
# Search 2: {agent_id}_messages (chat history)
|
||||
try:
|
||||
msg_filter = None
|
||||
if chat_id:
|
||||
# Payload schema differs across ingesters: some use chat_id, others channel_id.
|
||||
msg_filter = qmodels.Filter(
|
||||
should=[
|
||||
qmodels.FieldCondition(key="chat_id", match=qmodels.MatchValue(value=str(chat_id))),
|
||||
qmodels.FieldCondition(key="channel_id", match=qmodels.MatchValue(value=str(chat_id))),
|
||||
]
|
||||
)
|
||||
results = self.qdrant_client.search(
|
||||
collection_name=messages_collection,
|
||||
query_vector=embedding,
|
||||
query_filter=msg_filter,
|
||||
limit=limit,
|
||||
with_payload=True
|
||||
)
|
||||
|
||||
for r in results:
|
||||
if r.score > 0.4: # Higher threshold for messages
|
||||
# Higher threshold for messages; even higher when user asks about docs to avoid pulling old chatter.
|
||||
msg_thresh = 0.5 if is_doc_query else 0.4
|
||||
if r.score > msg_thresh:
|
||||
text = r.payload.get("text", r.payload.get("content", ""))
|
||||
# Skip very short or system messages
|
||||
if len(text) > 20 and not text.startswith("<"):
|
||||
if is_doc_query and topic_keywords:
|
||||
tl = text.lower()
|
||||
if not any(k in tl for k in topic_keywords):
|
||||
continue
|
||||
all_results.append({
|
||||
"text": text,
|
||||
"type": "message",
|
||||
@@ -446,18 +474,21 @@ class MemoryRetrieval:
|
||||
results = self.qdrant_client.search(
|
||||
collection_name=docs_collection,
|
||||
query_vector=embedding,
|
||||
limit=3, # Less docs, they're usually longer
|
||||
limit=6 if is_doc_query else 3, # Pull more docs for explicit doc queries
|
||||
with_payload=True
|
||||
)
|
||||
|
||||
for r in results:
|
||||
if r.score > 0.5: # Higher threshold for docs
|
||||
# When user asks about PDF/catalogs, relax threshold so docs show up more reliably.
|
||||
doc_thresh = 0.35 if is_doc_query else 0.5
|
||||
if r.score > doc_thresh:
|
||||
text = r.payload.get("text", r.payload.get("content", ""))
|
||||
if len(text) > 30:
|
||||
all_results.append({
|
||||
"text": text[:500], # Truncate long docs
|
||||
"type": "knowledge",
|
||||
"score": r.score,
|
||||
# Slightly boost docs for doc queries so they win vs chat snippets.
|
||||
"score": (r.score + 0.12) if is_doc_query else r.score,
|
||||
"source": "docs"
|
||||
})
|
||||
except Exception as e:
|
||||
@@ -614,7 +645,8 @@ class MemoryRetrieval:
|
||||
message_text: str,
|
||||
response_text: str,
|
||||
chat_id: str,
|
||||
message_type: str = "conversation"
|
||||
message_type: str = "conversation",
|
||||
metadata: Optional[Dict[str, Any]] = None,
|
||||
) -> bool:
|
||||
"""
|
||||
Store a message exchange in agent-specific Qdrant collection.
|
||||
@@ -656,23 +688,27 @@ class MemoryRetrieval:
|
||||
|
||||
# Store in Qdrant
|
||||
point_id = str(uuid.uuid4())
|
||||
payload = {
|
||||
"text": combined_text[:5000], # Limit payload size
|
||||
"user_message": message_text[:2000],
|
||||
"assistant_response": response_text[:3000],
|
||||
"user_id": user_id,
|
||||
"username": username,
|
||||
"chat_id": chat_id,
|
||||
"agent_id": agent_id,
|
||||
"type": message_type,
|
||||
"timestamp": datetime.utcnow().isoformat()
|
||||
}
|
||||
if metadata and isinstance(metadata, dict):
|
||||
payload["metadata"] = metadata
|
||||
|
||||
self.qdrant_client.upsert(
|
||||
collection_name=messages_collection,
|
||||
points=[
|
||||
qmodels.PointStruct(
|
||||
id=point_id,
|
||||
vector=embedding,
|
||||
payload={
|
||||
"text": combined_text[:5000], # Limit payload size
|
||||
"user_message": message_text[:2000],
|
||||
"assistant_response": response_text[:3000],
|
||||
"user_id": user_id,
|
||||
"username": username,
|
||||
"chat_id": chat_id,
|
||||
"agent_id": agent_id,
|
||||
"type": message_type,
|
||||
"timestamp": datetime.utcnow().isoformat()
|
||||
}
|
||||
payload=payload
|
||||
)
|
||||
]
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user