Fix pgvector integration: VECTOR(512) for voyage-3-lite, register_vector on connect

- voyage-3-lite returns 512 dims (not 1024) — migrated column + schema
- register_vector now called once at connection time, not per-query
- Removes per-function register_vector calls that caused type cast conflicts

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

Knowledgebase/schema.sql

@@ -2,6 +2,7 @@
 -- PostgreSQL, bez pgvector (embeddingy jako double precision[])
 -- pg_trgm pro fuzzy matching (volitelné)
 
+CREATE EXTENSION IF NOT EXISTS vector;
 CREATE EXTENSION IF NOT EXISTS pg_trgm;
 
 -- ─── Conversation sessions ────────────────────────────────────────────────────
@@ -56,7 +57,7 @@ CREATE TABLE IF NOT EXISTS kb_memories (
     -- search
     tags        TEXT[]          DEFAULT '{}',
     importance  FLOAT           DEFAULT 0.5,   -- 0..1
-    embedding   double precision[],  -- Voyage AI voyage-3-lite (1024-dim), Python-side similarity
+    embedding   vector(512),         -- Voyage AI voyage-3-lite (512-dim)
     fts         TSVECTOR,
 
     -- lifecycle
@@ -94,4 +95,5 @@ CREATE INDEX IF NOT EXISTS kb_memories_project_idx   ON kb_memories(project);
 CREATE INDEX IF NOT EXISTS kb_memories_importance_idx ON kb_memories(importance DESC);
 CREATE INDEX IF NOT EXISTS kb_memories_created_idx   ON kb_memories(created_at DESC);
 CREATE INDEX IF NOT EXISTS kb_memories_session_idx   ON kb_memories(session_id);
--- Note: embedding je double precision[] — similarity se počítá Python-side po FTS pre-filtru
+-- Vector index (aktivovat po ~1k řádcích pro lepší recall)
+-- CREATE INDEX kb_memories_vec_idx ON kb_memories USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100);

Knowledgebase/server.py

@@ -34,7 +34,8 @@ PG_PASSWORD = os.getenv("PG_PASSWORD", "Vlado7309208104++")
 PG_DB       = os.getenv("PG_DB",       "knowledgebase")
 
 VOYAGE_API_KEY = os.getenv("VOYAGE_API_KEY", "")
-EMBED_MODEL    = "voyage-3-lite"   # 1024-dim, fast & cheap
+EMBED_MODEL    = "voyage-3-lite"   # 512-dim, fast & cheap
+EMBED_DIM      = 512
 
 # ─── Logging ─────────────────────────────────────────────────────────────────
 
@@ -60,6 +61,12 @@ def get_conn() -> psycopg.Connection:
         row_factory=dict_row,
         autocommit=False,
     )
+    # register_vector jednou na connection — umožní psycopg správně serializovat np.array jako vector
+    try:
+        from pgvector.psycopg import register_vector
+        register_vector(_conn)
+    except Exception as e:
+        log(f"pgvector register warning: {e}")
     log(f"Connected to {PG_DB}@{PG_HOST}")
     return _conn
 
@@ -156,9 +163,7 @@ def store_memory(
     try:
         with conn.transaction():
             if embedding:
-                from pgvector.psycopg import register_vector
                 import numpy as np
-                register_vector(conn)
                 row = conn.execute(
                     """
                     INSERT INTO kb_memories
@@ -312,9 +317,7 @@ def _insert_memory_in_tx(conn, data: dict):
     """Helper: insert memory within an existing transaction."""
     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
@@ -420,9 +423,7 @@ def search(
         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] = []
@@ -700,9 +701,7 @@ 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(np.array(new_emb))
     if title is not None: