own-rag-cli 0.0.12-snapshot → 0.0.13-snapshot

package/README.md

@@ -1,4 +1,4 @@
-# MCP binary checksum (SHA-256, payload without shebang): `286e52abcbc295893b501fc8aae092420c452a4d3a4242e25c88e646398190b6`
+# MCP binary checksum (SHA-256, payload without shebang): `c74a07c0d8349fad54f3c5aa876296f9ed20239ffa0f2e6483e57796c7831e2a`
 
 # own-rag
 

package/bin/mcp_server.py

@@ -16,6 +16,8 @@ import json
 import logging
 import getpass
 import shutil
+import threading
+import time
 from collections.abc import Iterator
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import dataclass
@@ -392,6 +394,21 @@ MAX_FILE_SIZE_BYTES = 500 * 1024  # 500 KB
 TOP_K_RESULTS = 7
 MAX_QUERY_RESULTS = 30
 
+WARMUP_ENABLED = os.environ.get("MCP_WARMUP_ENABLED", "true").strip().lower() in {"1", "true", "yes", "on"}
+WARMUP_MODE = os.environ.get("MCP_WARMUP_MODE", "single").strip().lower()
+if WARMUP_MODE not in {"single", "ensemble"}:
+    WARMUP_MODE = "single"
+WARMUP_QUERY = os.environ.get("MCP_WARMUP_QUERY", "warmup semantic search")
+WARMUP_TOP_K = max(1, min(_parse_int(os.environ.get("MCP_WARMUP_TOP_K", "1"), 1, min_value=1), 20))
+
+KEEPALIVE_ENABLED = os.environ.get("MCP_KEEPALIVE_ENABLED", "true").strip().lower() in {"1", "true", "yes", "on"}
+KEEPALIVE_MODE = os.environ.get("MCP_KEEPALIVE_MODE", "single").strip().lower()
+if KEEPALIVE_MODE not in {"single", "ensemble"}:
+    KEEPALIVE_MODE = "single"
+KEEPALIVE_QUERY = os.environ.get("MCP_KEEPALIVE_QUERY", WARMUP_QUERY)
+KEEPALIVE_TOP_K = max(1, min(_parse_int(os.environ.get("MCP_KEEPALIVE_TOP_K", "1"), 1, min_value=1), 20))
+KEEPALIVE_INTERVAL_SEC = _parse_int(os.environ.get("MCP_KEEPALIVE_INTERVAL_SEC", "600"), 600, min_value=60)
+
 # Filtros de varredura
 IGNORED_DIRS = {
     ".git", "node_modules", "__pycache__", ".venv", "venv", "env",
@@ -1339,6 +1356,88 @@ def _run_ensemble_mode(query: str, top_k: int) -> tuple[list[FusedHit], list[str
     return reranked_hits, branch_errors, rerank_applied, rerank_error
 
 
+def _execute_background_search(mode: str, query: str, top_k: int, source: str) -> None:
+    started = time.perf_counter()
+    try:
+        if mode == "ensemble":
+            hits, branch_errors, rerank_applied, rerank_error = _run_ensemble_mode(query, top_k)
+        else:
+            hits, branch_errors, rerank_applied, rerank_error = _run_single_mode(query, top_k)
+
+        elapsed_ms = int((time.perf_counter() - started) * 1000)
+        log.info(
+            "%s concluido: mode=%s top_k=%s elapsed_ms=%s hits=%s branch_errors=%s rerank=%s",
+            source,
+            mode,
+            top_k,
+            elapsed_ms,
+            len(hits),
+            len(branch_errors),
+            rerank_applied,
+        )
+        _log_tool_usage(
+            event="maintenance_call",
+            tool_name=source,
+            details={
+                "status": "ok",
+                "mode": mode,
+                "query_len": len(query),
+                "top_k": top_k,
+                "elapsed_ms": elapsed_ms,
+                "result_count": len(hits),
+                "branch_errors": len(branch_errors),
+                "rerank_applied": rerank_applied,
+                "rerank_error": rerank_error,
+            },
+        )
+    except Exception as e:
+        elapsed_ms = int((time.perf_counter() - started) * 1000)
+        log.warning("%s falhou após %sms: %s", source, elapsed_ms, e)
+        _log_tool_usage(
+            event="maintenance_call",
+            tool_name=source,
+            details={
+                "status": "error",
+                "mode": mode,
+                "query_len": len(query),
+                "top_k": top_k,
+                "elapsed_ms": elapsed_ms,
+                "error": str(e),
+            },
+        )
+
+
+_keepalive_thread: threading.Thread | None = None
+_keepalive_stop_event = threading.Event()
+
+
+def _start_keepalive_thread() -> None:
+    global _keepalive_thread
+    if not KEEPALIVE_ENABLED:
+        log.info("Keepalive desabilitado (MCP_KEEPALIVE_ENABLED=false).")
+        return
+    if _keepalive_thread is not None and _keepalive_thread.is_alive():
+        return
+
+    def _runner() -> None:
+        log.info(
+            "Keepalive iniciado: intervalo=%ss mode=%s top_k=%s",
+            KEEPALIVE_INTERVAL_SEC,
+            KEEPALIVE_MODE,
+            KEEPALIVE_TOP_K,
+        )
+        while not _keepalive_stop_event.wait(KEEPALIVE_INTERVAL_SEC):
+            _execute_background_search(
+                mode=KEEPALIVE_MODE,
+                query=KEEPALIVE_QUERY,
+                top_k=KEEPALIVE_TOP_K,
+                source="semantic_search_keepalive",
+            )
+
+    _keepalive_thread = threading.Thread(target=_runner, name="rag-keepalive", daemon=True)
+    _keepalive_thread.start()
+
+
 # ---------------------------------------------------------------------------
 # Servidor MCP via FastMCP
 # ---------------------------------------------------------------------------
@@ -1799,6 +1898,16 @@ if __name__ == "__main__":
     log.info("Reranker: %s (enabled=%s, quant=%s)", RERANK_MODEL_ID, RERANK_ENABLED, RERANKER_QUANTIZATION)
     log.info("Pasta de modelos locais: %s", MODEL_DIR)
     log.info("Uso MCP será registrado em: %s", MCP_USAGE_LOG_PATH)
+    log.info(
+        "Warmup: enabled=%s mode=%s top_k=%s | Keepalive: enabled=%s interval=%ss mode=%s top_k=%s",
+        WARMUP_ENABLED,
+        WARMUP_MODE,
+        WARMUP_TOP_K,
+        KEEPALIVE_ENABLED,
+        KEEPALIVE_INTERVAL_SEC,
+        KEEPALIVE_MODE,
+        KEEPALIVE_TOP_K,
+    )
 
     # Pré-aquece somente conexão Chroma; modelos ficam lazy para poupar RAM.
     try:
@@ -1810,4 +1919,16 @@ if __name__ == "__main__":
         log.error("Falha ao inicializar ChromaDB: %s", e)
         log.error("O servidor continuará, mas as ferramentas retornarão erro até o ChromaDB estar disponível.")
 
+    if WARMUP_ENABLED:
+        _execute_background_search(
+            mode=WARMUP_MODE,
+            query=WARMUP_QUERY,
+            top_k=WARMUP_TOP_K,
+            source="semantic_search_warmup",
+        )
+    else:
+        log.info("Warmup desabilitado (MCP_WARMUP_ENABLED=false).")
+
+    _start_keepalive_thread()
+
     mcp.run(transport="stdio")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "own-rag-cli",
-  "version": "0.0.12-snapshot",
+  "version": "0.0.13-snapshot",
   "description": "Local RAG setup with ChromaDB + MCP server (Jina/BGE hybrid support).",
   "license": "MIT",
   "private": false,