janssen/Knowledgebase/server.py

#!/usr/bin/env python3
"""
Knowledgebase MCP server
========================
Persistentní paměť pro Claude konverzace a znalosti.

Vyhledávání:
  - Full-text (tsvector, vždy dostupné)
  - Sémantické (Voyage AI embeddingy + Python cosine similarity reranking)
  - Hybridní kombinace obou

Env proměnné:
  PG_HOST, PG_PORT, PG_USER, PG_PASSWORD, PG_DB
  VOYAGE_API_KEY   — pro vektorové embeddingy (volitelné)
"""

import json
import os
import sys
import traceback
from datetime import datetime
from typing import Any, Optional

import psycopg
from psycopg.rows import dict_row
from mcp.server.fastmcp import FastMCP

# ─── Config ──────────────────────────────────────────────────────────────────

PG_HOST     = os.getenv("PG_HOST",     "192.168.1.76")
PG_PORT     = int(os.getenv("PG_PORT", "5432"))
PG_USER     = os.getenv("PG_USER",     "vladimir.buzalka")
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"   # 512-dim, fast & cheap
EMBED_DIM      = 512

# ─── Logging ─────────────────────────────────────────────────────────────────

def log(msg: str):
    print(f"[KB] {msg}", file=sys.stderr, flush=True)

# ─── DB connection ───────────────────────────────────────────────────────────

_conn: Optional[psycopg.Connection] = None

def get_conn() -> psycopg.Connection:
    global _conn
    if _conn is not None and not _conn.closed:
        try:
            _conn.execute("SELECT 1")
            return _conn
        except Exception:
            pass
    _conn = psycopg.connect(
        host=PG_HOST, port=PG_PORT,
        user=PG_USER, password=PG_PASSWORD,
        dbname=PG_DB,
        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

# ─── Embeddings ──────────────────────────────────────────────────────────────

_voyage_client = None

def get_embedding(text: str) -> Optional[list[float]]:
    """Return 1024-dim embedding via Voyage AI, or None if unavailable."""
    global _voyage_client
    if not VOYAGE_API_KEY:
        return None
    try:
        if _voyage_client is None:
            import voyageai
            _voyage_client = voyageai.Client(api_key=VOYAGE_API_KEY)
        result = _voyage_client.embed([text[:8000]], model=EMBED_MODEL)
        return result.embeddings[0]
    except Exception as e:
        log(f"Embedding error: {e}")
        return None

# ─── Helpers ─────────────────────────────────────────────────────────────────

def _row_to_dict(row: dict) -> dict:
    """Serialize DB row for JSON output."""
    out = {}
    for k, v in row.items():
        if isinstance(v, datetime):
            out[k] = v.isoformat()
        elif isinstance(v, list) and k == "embedding":
            out[k] = None  # don't return raw vectors
        else:
            out[k] = v
    return out

def _fmt_memories(rows: list[dict]) -> str:
    if not rows:
        return "No results."
    parts = []
    for r in rows:
        score = f"  score={r.get('score', ''):.3f}" if r.get('score') is not None else ""
        parts.append(
            f"[{r['id']}] {r.get('mem_type','?').upper()} | {r.get('project') or '—'}"
            f"{score}\n"
            f"  Title: {r.get('title') or '—'}\n"
            f"  Tags:  {', '.join(r.get('tags') or []) or '—'}\n"
            f"  Date:  {r.get('created_at','')}\n"
            f"  Content:\n{_indent(r.get('content',''), 4)}"
        )
    return "\n\n".join(parts)

def _indent(text: str, n: int) -> str:
    pad = " " * n
    return "\n".join(pad + line for line in text.splitlines())

# ─── MCP server ──────────────────────────────────────────────────────────────

mcp = FastMCP("knowledgebase")

# ──────────────────────────────────────────────────────────────────────────────
# STORE MEMORY
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def store_memory(
    content: str,
    mem_type: str = "fact",
    title: Optional[str] = None,
    summary: Optional[str] = None,
    tags: Optional[list[str]] = None,
    project: Optional[str] = None,
    source: Optional[str] = None,
    session_id: Optional[str] = None,
    importance: float = 0.5,
    meta: Optional[dict] = None,
) -> str:
    """
    Uloží jedno paměťové záznam (fakt, rozhodnutí, preference, summary, …).

    mem_type: fact | decision | preference | summary | document | email | project | person | other
    importance: 0.0 (triviální) … 1.0 (kritické), default 0.5
    tags: seznam klíčových slov pro filtrování
    session_id: pokud pochází z konkrétní konverzace

    Příklad:
      store_memory("Vlado preferuje stručné odpovědi bez trailing summary",
                   mem_type="preference", tags=["komunikace", "styl"])
    """
    embedding = get_embedding(
        f"{title or ''} {summary or ''} {content}"
    )
    conn = get_conn()
    try:
        with conn.transaction():
            if embedding:
                import numpy as np
                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()
        log(traceback.format_exc())
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# STORE CONVERSATION
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def store_conversation(
    messages: list[dict],
    session_id: Optional[str] = None,
    title: Optional[str] = None,
    summary: Optional[str] = None,
    project: Optional[str] = None,
    tags: Optional[list[str]] = None,
    key_memories: Optional[list[dict]] = None,
) -> str:
    """
    Uloží celou konverzaci (seznam zpráv) jako session + automaticky
    extrahuje key_memories jako samostatné záznamy.

    messages: [{"role": "user"|"assistant", "content": "..."}]
    session_id: unikátní ID session (pokud není, vygeneruje se z timestampu)
    summary: shrnutí konverzace (doporučeno předat)
    key_memories: klíčové fakty/rozhodnutí z konverzace k samostatnému uložení
      [{"content": "...", "mem_type": "fact", "title": "...", "tags": [...], "importance": 0.7}]

    Příklad:
      store_conversation(
        messages=[...],
        session_id="2026-06-06-knowledgebase",
        title="Návrh Knowledgebase systému",
        summary="Vlado požadoval paměťový MCP server, rozhodli jsme se pro PG+pgvector+Voyage",
        project="knowledgebase",
        key_memories=[
          {"content": "Rozhodnutí: PostgreSQL + pgvector + Voyage AI embeddings",
           "mem_type": "decision", "importance": 0.9, "tags": ["architektura"]},
        ]
      )
    """
    import uuid
    sid = session_id or f"session-{datetime.utcnow().strftime('%Y%m%d-%H%M%S')}-{uuid.uuid4().hex[:6]}"

    conn = get_conn()
    try:
        with conn.transaction():
            # upsert session
            conn.execute(
                """
                INSERT INTO kb_sessions (id, title, summary, project, tags, msg_count, ended_at)
                VALUES (%s, %s, %s, %s, %s, %s, NOW())
                ON CONFLICT (id) DO UPDATE SET
                    title     = EXCLUDED.title,
                    summary   = EXCLUDED.summary,
                    project   = EXCLUDED.project,
                    tags      = EXCLUDED.tags,
                    msg_count = EXCLUDED.msg_count,
                    ended_at  = NOW()
                """,
                (sid, title, summary, project, tags or [], len(messages)),
            )

            # upsert messages
            conn.execute("DELETE FROM kb_messages WHERE session_id = %s", (sid,))
            for i, msg in enumerate(messages):
                conn.execute(
                    """
                    INSERT INTO kb_messages (session_id, role, content, seq)
                    VALUES (%s, %s, %s, %s)
                    """,
                    (sid, msg.get("role", "unknown"), msg.get("content", ""), i),
                )

            # summary jako paměťový záznam
            if summary:
                emb = get_embedding(f"{title or ''} {summary}")
                _insert_memory_in_tx(conn, {
                    "mem_type": "summary",
                    "title": title or f"Session {sid}",
                    "content": summary,
                    "session_id": sid,
                    "project": project,
                    "tags": tags or [],
                    "importance": 0.6,
                    "embedding": emb,
                })

            # key memories
            stored_km = 0
            for km in (key_memories or []):
                if not km.get("content"):
                    continue
                emb = get_embedding(
                    f"{km.get('title','') or ''} {km.get('content','')}"
                )
                _insert_memory_in_tx(conn, {
                    "mem_type":   km.get("mem_type", "fact"),
                    "title":      km.get("title"),
                    "content":    km["content"],
                    "summary":    km.get("summary"),
                    "session_id": sid,
                    "project":    km.get("project", project),
                    "tags":       km.get("tags", tags or []),
                    "importance": km.get("importance", 0.5),
                    "source":     km.get("source"),
                    "embedding":  emb,
                    "meta":       km.get("meta", {}),
                })
                stored_km += 1

        return (
            f"Stored session '{sid}' with {len(messages)} messages"
            f" + {stored_km} key memories."
        )
    except Exception as e:
        conn.rollback()
        log(traceback.format_exc())
        return f"Error: {e}"


def _insert_memory_in_tx(conn, data: dict):
    """Helper: insert memory within an existing transaction."""
    embedding = data.get("embedding")
    if embedding:
        import numpy as np
        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",{}))),
        )


# ──────────────────────────────────────────────────────────────────────────────
# SEARCH
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def search(
    query: str,
    types: Optional[list[str]] = None,
    project: Optional[str] = None,
    tags: Optional[list[str]] = None,
    limit: int = 10,
    min_importance: float = 0.0,
    include_sessions: bool = False,
) -> str:
    """
    Hybridní vyhledávání v paměti.
    Kombinuje full-text (vždy) + vektorové sémantické (pokud embeddingy dostupné).

    query: přirozený jazyk nebo klíčová slova
    types: ['fact','decision','preference','summary','document','email','project','person']
    project: filtrovat dle projektu
    tags: musí obsahovat alespoň jeden z těchto tagů
    limit: max počet výsledků (default 10)
    min_importance: min hodnota důležitosti 0..1
    include_sessions: zahrnout i výsledky z session summaries

    Příklad:
      search("PostgreSQL architektura", project="knowledgebase", types=["decision"])
    """
    conn = get_conn()
    results = []

    try:
        # ── Full-text search ──
        conditions = ["deleted = FALSE", "fts @@ plainto_tsquery('simple', %s)"]
        params: list[Any] = [query]

        if types:
            conditions.append(f"mem_type = ANY(%s)")
            params.append(types)
        if project:
            conditions.append("project = %s")
            params.append(project)
        if tags:
            conditions.append("tags && %s")
            params.append(tags)
        if min_importance > 0:
            conditions.append("importance >= %s")
            params.append(min_importance)
        if not include_sessions:
            conditions.append("mem_type != 'summary' OR session_id IS NOT NULL")

        where = " AND ".join(conditions)
        rows = conn.execute(
            f"""
            SELECT id, mem_type, title, content, summary, tags,
                   project, source, session_id, importance, created_at,
                   ts_rank(fts, plainto_tsquery('simple', %s)) AS score
            FROM kb_memories
            WHERE {where}
            ORDER BY score DESC, importance DESC
            LIMIT %s
            """,
            [query] + params + [limit],
        ).fetchall()

        fts_ids = {r["id"] for r in rows}
        results = [_row_to_dict(r) for r in rows]

        # ── Vector search (nativní pgvector, <=> cosine distance) ──
        query_emb = get_embedding(query)
        if query_emb:
            try:
                import numpy as np

                vec_conditions = ["deleted = FALSE", "embedding IS NOT NULL"]
                vec_params2: list[Any] = []

                if types:
                    vec_conditions.append("mem_type = ANY(%s)")
                    vec_params2.append(types)
                if project:
                    vec_conditions.append("project = %s")
                    vec_params2.append(project)
                if tags:
                    vec_conditions.append("tags && %s")
                    vec_params2.append(tags)
                if min_importance > 0:
                    vec_conditions.append("importance >= %s")
                    vec_params2.append(min_importance)

                vec_where = " AND ".join(vec_conditions)
                qv = np.array(query_emb)
                # Pořadí %s musí odpovídat pořadí v SQL:
                # 1. WHERE podmínky (vec_params2)
                # 2. SELECT score: embedding <=> %s
                # 3. ORDER BY: embedding <=> %s
                # 4. LIMIT %s
                vec_rows = conn.execute(
                    f"""
                    SELECT id, mem_type, title, content, summary, tags,
                           project, source, session_id, importance, created_at,
                           1 - (embedding <=> %s) AS score
                    FROM kb_memories
                    WHERE {vec_where}
                    ORDER BY embedding <=> %s
                    LIMIT %s
                    """,
                    [qv] + vec_params2 + [qv, limit],
                ).fetchall()

                for r in vec_rows:
                    if r["id"] not in fts_ids:
                        results.append(_row_to_dict(r))

            except Exception as e:
                log(f"Vector search error: {e}")

        # deduplicate & sort by score
        seen = set()
        deduped = []
        for r in results:
            if r["id"] not in seen:
                seen.add(r["id"])
                deduped.append(r)

        deduped.sort(key=lambda x: (x.get("score") or 0, x.get("importance", 0)), reverse=True)

        return _fmt_memories(deduped[:limit])

    except Exception as e:
        log(traceback.format_exc())
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# GET CONTEXT (kontext pro aktuální konverzaci)
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def get_context(
    topic: str,
    project: Optional[str] = None,
    limit: int = 8,
    include_preferences: bool = True,
) -> str:
    """
    Vrátí nejrelevantnější paměti pro daný kontext/téma.
    Automaticky přidá preference uživatele (pokud exist.)

    Použití na začátku konverzace:
      get_context("IWRS pacienti, notifikace, MongoDB")
    """
    conn = get_conn()
    parts = []

    # ── Main context ──
    main_result = search(
        query=topic,
        project=project,
        limit=limit,
        min_importance=0.3,
    )
    parts.append("=== RELEVANTNÍ PAMĚTI ===\n" + main_result)

    # ── User preferences ──
    if include_preferences:
        try:
            rows = conn.execute(
                """
                SELECT id, mem_type, title, content, tags, importance, created_at
                FROM kb_memories
                WHERE mem_type = 'preference'
                  AND deleted = FALSE
                ORDER BY importance DESC, created_at DESC
                LIMIT 5
                """,
            ).fetchall()
            if rows:
                pref_lines = [_fmt_memories([_row_to_dict(r) for r in rows])]
                parts.append("=== PREFERENCE UŽIVATELE ===\n" + "\n".join(pref_lines))
        except Exception as e:
            log(f"Preferences error: {e}")

    return "\n\n".join(parts)


# ──────────────────────────────────────────────────────────────────────────────
# GET RECENT
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def get_recent(
    limit: int = 10,
    mem_type: Optional[str] = None,
    project: Optional[str] = None,
) -> str:
    """
    Vrátí nejnovější paměti, volitelně filtrované.

    Příklad:
      get_recent(limit=5, mem_type="decision")
    """
    conn = get_conn()
    conditions = ["deleted = FALSE"]
    params: list[Any] = []

    if mem_type:
        conditions.append("mem_type = %s")
        params.append(mem_type)
    if project:
        conditions.append("project = %s")
        params.append(project)

    where = " AND ".join(conditions)
    try:
        rows = conn.execute(
            f"""
            SELECT id, mem_type, title, content, tags, project, importance, created_at
            FROM kb_memories
            WHERE {where}
            ORDER BY created_at DESC
            LIMIT %s
            """,
            params + [limit],
        ).fetchall()
        return _fmt_memories([_row_to_dict(r) for r in rows])
    except Exception as e:
        log(traceback.format_exc())
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# LIST SESSIONS
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def list_sessions(
    limit: int = 20,
    project: Optional[str] = None,
) -> str:
    """
    Vypíše přehled uložených konverzačních sessions.
    """
    conn = get_conn()
    conditions = []
    params: list[Any] = []
    if project:
        conditions.append("project = %s")
        params.append(project)

    where = "WHERE " + " AND ".join(conditions) if conditions else ""
    try:
        rows = conn.execute(
            f"""
            SELECT id, title, summary, project, tags, msg_count, started_at, ended_at
            FROM kb_sessions
            {where}
            ORDER BY ended_at DESC NULLS LAST
            LIMIT %s
            """,
            params + [limit],
        ).fetchall()
        if not rows:
            return "No sessions found."
        lines = []
        for r in rows:
            lines.append(
                f"[{r['id']}] {r.get('title') or '—'} | {r.get('project') or '—'}"
                f" | {r['msg_count']} msgs | {r.get('ended_at','')}\n"
                f"  {(r.get('summary') or '')[:120]}"
            )
        return "\n\n".join(lines)
    except Exception as e:
        log(traceback.format_exc())
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# GET SESSION
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def get_session(
    session_id: str,
    include_messages: bool = True,
    messages_limit: int = 100,
) -> str:
    """
    Vrátí detail konkrétní session (metadata + zprávy).
    """
    conn = get_conn()
    try:
        sess = conn.execute(
            "SELECT * FROM kb_sessions WHERE id = %s", (session_id,)
        ).fetchone()
        if not sess:
            return f"Session '{session_id}' not found."

        out = [
            f"Session: {sess['id']}",
            f"Title:   {sess.get('title') or '—'}",
            f"Project: {sess.get('project') or '—'}",
            f"Tags:    {', '.join(sess.get('tags') or [])}",
            f"Date:    {sess.get('ended_at') or sess.get('started_at')}",
            f"Summary:\n{_indent(sess.get('summary') or '—', 2)}",
        ]

        if include_messages:
            msgs = conn.execute(
                """
                SELECT role, content, seq FROM kb_messages
                WHERE session_id = %s ORDER BY seq LIMIT %s
                """,
                (session_id, messages_limit),
            ).fetchall()
            out.append(f"\n--- Messages ({len(msgs)}) ---")
            for m in msgs:
                role = m["role"].upper()
                content = m["content"]
                if len(content) > 500:
                    content = content[:500] + "…"
                out.append(f"\n[{role}]\n{content}")

        return "\n".join(out)
    except Exception as e:
        log(traceback.format_exc())
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# UPDATE MEMORY
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def update_memory(
    memory_id: int,
    content: Optional[str] = None,
    title: Optional[str] = None,
    summary: Optional[str] = None,
    tags: Optional[list[str]] = None,
    importance: Optional[float] = None,
    project: Optional[str] = None,
) -> str:
    """
    Aktualizuje existující paměťový záznam.
    Předej jen pole, která chceš změnit.
    """
    conn = get_conn()
    updates = []
    params: list[Any] = []

    if content is not None:
        updates.append("content = %s")
        params.append(content)
        new_emb = get_embedding(f"{title or ''} {content}")
        if new_emb:
            import numpy as np
            updates.append("embedding = %s")
            params.append(np.array(new_emb))
    if title is not None:
        updates.append("title = %s")
        params.append(title)
    if summary is not None:
        updates.append("summary = %s")
        params.append(summary)
    if tags is not None:
        updates.append("tags = %s")
        params.append(tags)
    if importance is not None:
        updates.append("importance = %s")
        params.append(importance)
    if project is not None:
        updates.append("project = %s")
        params.append(project)

    if not updates:
        return "Nothing to update."

    params.append(memory_id)
    try:
        with conn.transaction():
            conn.execute(
                f"UPDATE kb_memories SET {', '.join(updates)} WHERE id = %s",
                params,
            )
        return f"Memory {memory_id} updated."
    except Exception as e:
        conn.rollback()
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# DELETE MEMORY
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def delete_memory(memory_id: int, hard: bool = False) -> str:
    """
    Soft-delete (default) nebo hard-delete paměti.
    Soft: záznám zůstane v DB, jen se skryje z výsledků.
    Hard: smaže fyzicky.
    """
    conn = get_conn()
    try:
        with conn.transaction():
            if hard:
                conn.execute("DELETE FROM kb_memories WHERE id = %s", (memory_id,))
                return f"Memory {memory_id} permanently deleted."
            else:
                conn.execute(
                    "UPDATE kb_memories SET deleted = TRUE WHERE id = %s", (memory_id,)
                )
                return f"Memory {memory_id} soft-deleted."
    except Exception as e:
        conn.rollback()
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────
# STATS
# ──────────────────────────────────────────────────────────────────────────────

@mcp.tool()
def stats() -> str:
    """
    Přehled obsahu databáze: počty záznamů dle typu, sessions, projekty.
    """
    conn = get_conn()
    try:
        type_counts = conn.execute(
            """
            SELECT mem_type, COUNT(*) AS cnt
            FROM kb_memories
            WHERE deleted = FALSE
            GROUP BY mem_type ORDER BY cnt DESC
            """
        ).fetchall()

        session_count = conn.execute(
            "SELECT COUNT(*) AS cnt FROM kb_sessions"
        ).fetchone()["cnt"]

        projects = conn.execute(
            """
            SELECT project, COUNT(*) AS cnt
            FROM kb_memories
            WHERE deleted = FALSE AND project IS NOT NULL
            GROUP BY project ORDER BY cnt DESC LIMIT 10
            """
        ).fetchall()

        embed_count = conn.execute(
            "SELECT COUNT(*) AS cnt FROM kb_memories WHERE embedding IS NOT NULL"
        ).fetchone()["cnt"]

        lines = ["=== Knowledgebase Stats ==="]
        lines.append(f"\nSessions: {session_count}")
        lines.append(f"Embeddings: {embed_count} (Voyage AI {'active' if VOYAGE_API_KEY else 'not configured'})")

        lines.append("\nMemories by type:")
        for r in type_counts:
            lines.append(f"  {r['mem_type']:15} {r['cnt']:>5}")

        if projects:
            lines.append("\nTop projects:")
            for r in projects:
                lines.append(f"  {r['project']:25} {r['cnt']:>5}")

        return "\n".join(lines)
    except Exception as e:
        log(traceback.format_exc())
        return f"Error: {e}"


# ──────────────────────────────────────────────────────────────────────────────

if __name__ == "__main__":
    log("Starting Knowledgebase MCP server...")
    log(f"  DB: {PG_DB}@{PG_HOST}:{PG_PORT}")
    log(f"  Embeddings: {'Voyage AI' if VOYAGE_API_KEY else 'disabled (set VOYAGE_API_KEY)'}")
    mcp.run(transport="stdio")