Upgrade Knowledgebase to native pgvector (VECTOR type + ivfflat index)

- Migrated embedding column from double precision[] to VECTOR(1024)
- Now uses native <=> cosine operator for SQL-level vector search
- Added pgvector to requirements
- Fixed collation mismatch on all DBs after pgvector/pgvector:pg18 image swap

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

Knowledgebase/server.py

@@ -15,7 +15,6 @@ Env proměnné:
 """
 
 import json
-import math
 import os
 import sys
 import traceback
@@ -117,14 +116,6 @@ def _indent(text: str, n: int) -> str:
     pad = " " * n
     return "\n".join(pad + line for line in text.splitlines())
 
-def _cosine(a: list[float], b: list[float]) -> float:
-    dot = sum(x * y for x, y in zip(a, b))
-    na  = math.sqrt(sum(x * x for x in a))
-    nb  = math.sqrt(sum(x * x for x in b))
-    if na == 0 or nb == 0:
-        return 0.0
-    return dot / (na * nb)
-
 # ─── MCP server ──────────────────────────────────────────────────────────────
 
 mcp = FastMCP("knowledgebase")
@@ -164,18 +155,35 @@ def store_memory(
     conn = get_conn()
     try:
         with conn.transaction():
-            row = conn.execute(
-                """
-                INSERT INTO kb_memories
-                    (mem_type, title, content, summary, tags, project,
-                     source, session_id, importance, embedding, meta)
-                VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)
-                RETURNING id, created_at
-                """,
-                (mem_type, title, content, summary,
-                 tags or [], project, source, session_id,
-                 importance, embedding, json.dumps(meta or {})),
-            ).fetchone()
+            if embedding:
+                from pgvector.psycopg import register_vector
+                import numpy as np
+                register_vector(conn)
+                row = conn.execute(
+                    """
+                    INSERT INTO kb_memories
+                        (mem_type, title, content, summary, tags, project,
+                         source, session_id, importance, embedding, meta)
+                    VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)
+                    RETURNING id, created_at
+                    """,
+                    (mem_type, title, content, summary,
+                     tags or [], project, source, session_id,
+                     importance, np.array(embedding), json.dumps(meta or {})),
+                ).fetchone()
+            else:
+                row = conn.execute(
+                    """
+                    INSERT INTO kb_memories
+                        (mem_type, title, content, summary, tags, project,
+                         source, session_id, importance, meta)
+                    VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)
+                    RETURNING id, created_at
+                    """,
+                    (mem_type, title, content, summary,
+                     tags or [], project, source, session_id,
+                     importance, json.dumps(meta or {})),
+                ).fetchone()
         return f"Stored memory id={row['id']} at {row['created_at']}"
     except Exception as e:
         conn.rollback()
@@ -302,21 +310,41 @@ def store_conversation(
 
 def _insert_memory_in_tx(conn, data: dict):
     """Helper: insert memory within an existing transaction."""
-    conn.execute(
-        """
-        INSERT INTO kb_memories
-            (mem_type, title, content, summary, tags, project,
-             source, session_id, importance, embedding, meta)
-        VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)
-        """,
-        (data.get("mem_type","fact"), data.get("title"),
-         data["content"], data.get("summary"),
-         data.get("tags",[]), data.get("project"),
-         data.get("source"), data.get("session_id"),
-         data.get("importance",0.5),
-         data.get("embedding"),
-         json.dumps(data.get("meta",{}))),
-    )
+    embedding = data.get("embedding")
+    if embedding:
+        from pgvector.psycopg import register_vector
+        import numpy as np
+        register_vector(conn)
+        conn.execute(
+            """
+            INSERT INTO kb_memories
+                (mem_type, title, content, summary, tags, project,
+                 source, session_id, importance, embedding, meta)
+            VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)
+            """,
+            (data.get("mem_type","fact"), data.get("title"),
+             data["content"], data.get("summary"),
+             data.get("tags",[]), data.get("project"),
+             data.get("source"), data.get("session_id"),
+             data.get("importance",0.5),
+             np.array(embedding),
+             json.dumps(data.get("meta",{}))),
+        )
+    else:
+        conn.execute(
+            """
+            INSERT INTO kb_memories
+                (mem_type, title, content, summary, tags, project,
+                 source, session_id, importance, meta)
+            VALUES (%s,%s,%s,%s,%s,%s,%s,%s,%s,%s)
+            """,
+            (data.get("mem_type","fact"), data.get("title"),
+             data["content"], data.get("summary"),
+             data.get("tags",[]), data.get("project"),
+             data.get("source"), data.get("session_id"),
+             data.get("importance",0.5),
+             json.dumps(data.get("meta",{}))),
+        )
 
 
 # ──────────────────────────────────────────────────────────────────────────────
@@ -388,11 +416,14 @@ def search(
         fts_ids = {r["id"] for r in rows}
         results = [_row_to_dict(r) for r in rows]
 
-        # ── Vector reranking (Python-side cosine similarity) ──
-        # Fetch candidates with embeddings, compute cosine similarity, merge
+        # ── Vector search (nativní pgvector, <=> cosine distance) ──
         query_emb = get_embedding(query)
         if query_emb:
             try:
+                from pgvector.psycopg import register_vector
+                import numpy as np
+                register_vector(conn)
+
                 vec_conditions = ["deleted = FALSE", "embedding IS NOT NULL"]
                 vec_params2: list[Any] = []
 
@@ -414,24 +445,21 @@ def search(
                     f"""
                     SELECT id, mem_type, title, content, summary, tags,
                            project, source, session_id, importance, created_at,
-                           embedding
+                           1 - (embedding <=> %s::vector) AS score
                     FROM kb_memories
                     WHERE {vec_where}
-                    LIMIT 200
+                    ORDER BY embedding <=> %s::vector
+                    LIMIT %s
                     """,
-                    vec_params2,
+                    [np.array(query_emb), np.array(query_emb)] + vec_params2 + [limit],
                 ).fetchall()
 
                 for r in vec_rows:
-                    if r["id"] not in fts_ids and r["embedding"]:
-                        sim = _cosine(query_emb, r["embedding"])
-                        if sim > 0.5:  # threshold
-                            d = _row_to_dict(r)
-                            d["score"] = sim
-                            results.append(d)
+                    if r["id"] not in fts_ids:
+                        results.append(_row_to_dict(r))
 
             except Exception as e:
-                log(f"Vector reranking error: {e}")
+                log(f"Vector search error: {e}")
 
         # deduplicate & sort by score
         seen = set()
@@ -672,8 +700,11 @@ def update_memory(
         params.append(content)
         new_emb = get_embedding(f"{title or ''} {content}")
         if new_emb:
+            from pgvector.psycopg import register_vector
+            import numpy as np
+            register_vector(conn)
             updates.append("embedding = %s")
-            params.append(new_emb)
+            params.append(np.array(new_emb))
     if title is not None:
         updates.append("title = %s")
         params.append(title)