feat(MAT-13): Add conversation chunk RAG for Matrix chat history

Add semantic search over past conversations alongside existing memory facts.
New conversation_chunks table stores user-assistant exchanges with LLM-generated
summaries embedded for retrieval. Bot queries chunks on each message and injects
relevant past conversations into the system prompt. New exchanges are indexed
automatically after each bot response.

Memory-service: /chunks/store, /chunks/query, /chunks/bulk-store endpoints
Bot: chunk query + formatting, live indexing via asyncio.gather with memory extraction

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

bot.py

@@ -457,6 +457,39 @@ class MemoryClient:
             logger.warning("Memory list failed", exc_info=True)
             return []
 
+    async def store_chunk(self, user_id: str, room_id: str, chunk_text: str,
+                          summary: str, source_event_id: str = "", original_ts: float = 0.0):
+        if not self.enabled:
+            return
+        try:
+            async with httpx.AsyncClient(timeout=15.0) as client:
+                await client.post(
+                    f"{self.base_url}/chunks/store",
+                    json={
+                        "user_id": user_id, "room_id": room_id,
+                        "chunk_text": chunk_text, "summary": summary,
+                        "source_event_id": source_event_id, "original_ts": original_ts,
+                    },
+                )
+        except Exception:
+            logger.warning("Chunk store failed", exc_info=True)
+
+    async def query_chunks(self, query: str, user_id: str = "", room_id: str = "",
+                           top_k: int = 5) -> list[dict]:
+        if not self.enabled:
+            return []
+        try:
+            async with httpx.AsyncClient(timeout=10.0) as client:
+                resp = await client.post(
+                    f"{self.base_url}/chunks/query",
+                    json={"user_id": user_id, "room_id": room_id, "query": query, "top_k": top_k},
+                )
+                resp.raise_for_status()
+                return resp.json().get("results", [])
+        except Exception:
+            logger.warning("Chunk query failed", exc_info=True)
+            return []
+
 
 class AtlassianClient:
     """Fetches per-user Atlassian tokens from the portal and calls Atlassian REST APIs."""
@@ -1173,6 +1206,52 @@ class Bot:
         facts = [m["fact"] for m in memories]
         return "You have these memories about this user:\n" + "\n".join(f"- {f}" for f in facts)
 
+    @staticmethod
+    def _format_chunks(chunks: list[dict]) -> str:
+        """Format conversation chunk results as a system prompt section."""
+        if not chunks:
+            return ""
+        parts = ["Relevant past conversations:"]
+        for c in chunks:
+            ts = c.get("original_ts", 0)
+            date_str = time.strftime("%Y-%m-%d", time.gmtime(ts)) if ts else "unknown"
+            summary = c.get("summary", "")
+            text = c.get("chunk_text", "")
+            # Truncate chunk text to ~500 chars for context window efficiency
+            if len(text) > 500:
+                text = text[:500] + "..."
+            parts.append(f"\n### {summary} ({date_str})\n{text}")
+        return "\n".join(parts)
+
+    async def _store_conversation_chunk(self, user_message: str, ai_reply: str,
+                                          sender: str, room_id: str):
+        """Store a user-assistant exchange as a conversation chunk for RAG."""
+        if not self.llm or not self.memory.enabled:
+            return
+        chunk_text = f"User: {user_message}\nAssistant: {ai_reply}"
+        try:
+            resp = await self.llm.chat.completions.create(
+                model="claude-haiku",
+                messages=[
+                    {"role": "system", "content": (
+                        "Summarize this conversation exchange in 1-2 sentences for search indexing. "
+                        "Focus on the topic and key information discussed. Be concise. "
+                        "Write the summary in the same language as the conversation."
+                    )},
+                    {"role": "user", "content": chunk_text[:2000]},
+                ],
+                max_tokens=100,
+            )
+            summary = resp.choices[0].message.content.strip()
+        except Exception:
+            logger.warning("Chunk summarization failed, using truncated message", exc_info=True)
+            summary = user_message[:200]
+
+        await self.memory.store_chunk(
+            user_id=sender, room_id=room_id, chunk_text=chunk_text,
+            summary=summary, original_ts=time.time(),
+        )
+
     async def _extract_and_store_memories(self, user_message: str, ai_reply: str,
                                           existing_facts: list[str], model: str,
                                           sender: str, room_id: str):
@@ -2092,6 +2171,10 @@ class Bot:
         memories = await self.memory.query(sender, user_message, top_k=10) if sender else []
         memory_context = self._format_memories(memories)
 
+        # Query relevant conversation chunks (RAG over chat history)
+        chunks = await self.memory.query_chunks(search_query, user_id=sender or "", top_k=5)
+        chunk_context = self._format_chunks(chunks)
+
         # Include room document context (PDFs, Confluence pages, images uploaded to room)
         room_doc_context = ""
         room_docs = [e for e in self._room_document_context.get(room.room_id, [])
@@ -2114,6 +2197,8 @@ class Bot:
         messages = [{"role": "system", "content": SYSTEM_PROMPT}]
         if memory_context:
             messages.append({"role": "system", "content": memory_context})
+        if chunk_context:
+            messages.append({"role": "system", "content": chunk_context})
         if doc_context:
             messages.append({"role": "system", "content": doc_context})
         if room_doc_context:
@@ -2181,20 +2266,25 @@ class Bot:
             if reply:
                 await self._send_text(room.room_id, reply)
 
-            # Extract and store new memories (after reply sent, with timeout)
+            # Extract and store new memories + conversation chunk (after reply sent)
             if sender and reply:
                 existing_facts = [m["fact"] for m in memories]
                 try:
                     await asyncio.wait_for(
+                        asyncio.gather(
                             self._extract_and_store_memories(
                                 user_message, reply, existing_facts, model, sender, room.room_id
                             ),
-                        timeout=15.0,
+                            self._store_conversation_chunk(
+                                user_message, reply, sender, room.room_id
+                            ),
+                        ),
+                        timeout=20.0,
                     )
                 except asyncio.TimeoutError:
-                    logger.warning("Memory extraction timed out for %s", sender)
+                    logger.warning("Memory/chunk extraction timed out for %s", sender)
                 except Exception:
-                    logger.warning("Memory save failed", exc_info=True)
+                    logger.warning("Memory/chunk save failed", exc_info=True)
 
             # Auto-rename: only for group rooms with explicit opt-in (not DMs)
             if room.room_id in self.auto_rename_rooms:

memory-service/main.py

@@ -33,6 +33,26 @@ class QueryRequest(BaseModel):
     top_k: int = 10
 
 
+class ChunkStoreRequest(BaseModel):
+    user_id: str
+    room_id: str
+    chunk_text: str
+    summary: str
+    source_event_id: str = ""
+    original_ts: float = 0.0
+
+
+class ChunkQueryRequest(BaseModel):
+    user_id: str = ""
+    room_id: str = ""
+    query: str
+    top_k: int = 5
+
+
+class ChunkBulkStoreRequest(BaseModel):
+    chunks: list[ChunkStoreRequest]
+
+
 async def _embed(text: str) -> list[float]:
     """Get embedding vector from LiteLLM /embeddings endpoint."""
     async with httpx.AsyncClient(timeout=30.0) as client:
@@ -45,6 +65,19 @@ async def _embed(text: str) -> list[float]:
         return resp.json()["data"][0]["embedding"]
 
 
+async def _embed_batch(texts: list[str]) -> list[list[float]]:
+    """Get embedding vectors for a batch of texts."""
+    async with httpx.AsyncClient(timeout=60.0) as client:
+        resp = await client.post(
+            f"{LITELLM_URL}/embeddings",
+            json={"model": EMBED_MODEL, "input": texts},
+            headers={"Authorization": f"Bearer {LITELLM_KEY}"},
+        )
+        resp.raise_for_status()
+        data = resp.json()["data"]
+        return [item["embedding"] for item in sorted(data, key=lambda x: x["index"])]
+
+
 async def _init_db():
     """Create pgvector extension and memories table if not exists."""
     global pool
@@ -69,6 +102,25 @@ async def _init_db():
             ON memories USING ivfflat (embedding vector_cosine_ops)
             WITH (lists = 10)
         """)
+        # Conversation chunks table for RAG over chat history
+        await conn.execute(f"""
+            CREATE TABLE IF NOT EXISTS conversation_chunks (
+                id BIGSERIAL PRIMARY KEY,
+                user_id TEXT NOT NULL,
+                room_id TEXT NOT NULL,
+                chunk_text TEXT NOT NULL,
+                summary TEXT NOT NULL,
+                source_event_id TEXT DEFAULT '',
+                original_ts DOUBLE PRECISION NOT NULL,
+                embedding vector({EMBED_DIMS}) NOT NULL
+            )
+        """)
+        await conn.execute("""
+            CREATE INDEX IF NOT EXISTS idx_chunks_user_id ON conversation_chunks (user_id)
+        """)
+        await conn.execute("""
+            CREATE INDEX IF NOT EXISTS idx_chunks_room_id ON conversation_chunks (room_id)
+        """)
     logger.info("Database initialized (dims=%d)", EMBED_DIMS)
 
 
@@ -87,8 +139,9 @@ async def shutdown():
 async def health():
     if pool:
         async with pool.acquire() as conn:
-            count = await conn.fetchval("SELECT count(*) FROM memories")
+            mem_count = await conn.fetchval("SELECT count(*) FROM memories")
-        return {"status": "ok", "total_memories": count}
+            chunk_count = await conn.fetchval("SELECT count(*) FROM conversation_chunks")
+        return {"status": "ok", "total_memories": mem_count, "total_chunks": chunk_count}
     return {"status": "no_db"}
 
 
@@ -189,3 +242,129 @@ async def list_user_memories(user_id: str):
             for r in rows
         ],
     }
+
+
+# --- Conversation Chunks ---
+
+
+@app.post("/chunks/store")
+async def store_chunk(req: ChunkStoreRequest):
+    """Store a conversation chunk with its summary embedding."""
+    if not req.summary.strip():
+        raise HTTPException(400, "Empty summary")
+
+    embedding = await _embed(req.summary)
+    vec_literal = "[" + ",".join(str(v) for v in embedding) + "]"
+    ts = req.original_ts or time.time()
+
+    async with pool.acquire() as conn:
+        await conn.execute(
+            """
+            INSERT INTO conversation_chunks
+                (user_id, room_id, chunk_text, summary, source_event_id, original_ts, embedding)
+            VALUES ($1, $2, $3, $4, $5, $6, $7::vector)
+            """,
+            req.user_id, req.room_id, req.chunk_text, req.summary,
+            req.source_event_id, ts, vec_literal,
+        )
+    logger.info("Stored chunk for %s in %s: %s", req.user_id, req.room_id, req.summary[:60])
+    return {"stored": True}
+
+
+@app.post("/chunks/query")
+async def query_chunks(req: ChunkQueryRequest):
+    """Semantic search over conversation chunks. Filter by user_id and/or room_id."""
+    embedding = await _embed(req.query)
+    vec_literal = "[" + ",".join(str(v) for v in embedding) + "]"
+
+    # Build WHERE clause dynamically based on provided filters
+    conditions = []
+    params: list = [vec_literal]
+    idx = 2
+
+    if req.user_id:
+        conditions.append(f"user_id = ${idx}")
+        params.append(req.user_id)
+        idx += 1
+    if req.room_id:
+        conditions.append(f"room_id = ${idx}")
+        params.append(req.room_id)
+        idx += 1
+
+    where = f"WHERE {' AND '.join(conditions)}" if conditions else ""
+    params.append(req.top_k)
+
+    async with pool.acquire() as conn:
+        rows = await conn.fetch(
+            f"""
+            SELECT chunk_text, summary, room_id, user_id, original_ts, source_event_id,
+                   1 - (embedding <=> $1::vector) AS similarity
+            FROM conversation_chunks
+            {where}
+            ORDER BY embedding <=> $1::vector
+            LIMIT ${idx}
+            """,
+            *params,
+        )
+
+    results = [
+        {
+            "chunk_text": r["chunk_text"],
+            "summary": r["summary"],
+            "room_id": r["room_id"],
+            "user_id": r["user_id"],
+            "original_ts": r["original_ts"],
+            "source_event_id": r["source_event_id"],
+            "similarity": float(r["similarity"]),
+        }
+        for r in rows
+    ]
+    return {"results": results}
+
+
+@app.post("/chunks/bulk-store")
+async def bulk_store_chunks(req: ChunkBulkStoreRequest):
+    """Batch store conversation chunks. Embeds summaries in batches of 20."""
+    if not req.chunks:
+        return {"stored": 0}
+
+    stored = 0
+    batch_size = 20
+
+    for i in range(0, len(req.chunks), batch_size):
+        batch = req.chunks[i:i + batch_size]
+        summaries = [c.summary.strip() for c in batch]
+
+        try:
+            embeddings = await _embed_batch(summaries)
+        except Exception:
+            logger.error("Batch embed failed for chunks %d-%d", i, i + len(batch), exc_info=True)
+            continue
+
+        async with pool.acquire() as conn:
+            for chunk, embedding in zip(batch, embeddings):
+                vec_literal = "[" + ",".join(str(v) for v in embedding) + "]"
+                ts = chunk.original_ts or time.time()
+                await conn.execute(
+                    """
+                    INSERT INTO conversation_chunks
+                        (user_id, room_id, chunk_text, summary, source_event_id, original_ts, embedding)
+                    VALUES ($1, $2, $3, $4, $5, $6, $7::vector)
+                    """,
+                    chunk.user_id, chunk.room_id, chunk.chunk_text, chunk.summary,
+                    chunk.source_event_id, ts, vec_literal,
+                )
+                stored += 1
+
+    logger.info("Bulk stored %d chunks", stored)
+    return {"stored": stored}
+
+
+@app.get("/chunks/{user_id}/count")
+async def count_user_chunks(user_id: str):
+    """Count conversation chunks for a user."""
+    async with pool.acquire() as conn:
+        count = await conn.fetchval(
+            "SELECT count(*) FROM conversation_chunks WHERE user_id = $1", user_id,
+        )
+    return {"user_id": user_id, "count": count}