Add quickstart and RAG MCP starter docs.

Made-with: Cursor

docs/rag-mcp-starter/mcp_server.py

@@ -0,0 +1,85 @@
+#!/usr/bin/env python3
+import os
+
+import requests
+from fastmcp import FastMCP
+from qdrant_client import QdrantClient
+
+MCP_NAME = os.getenv("MCP_NAME", "local-rag")
+QDRANT_URL = os.getenv("QDRANT_URL", "http://127.0.0.1:6333")
+COLLECTION = os.getenv("QDRANT_COLLECTION", "docs")
+EMBED_MODEL = os.getenv("EMBED_MODEL", "nomic-embed-text")
+GEN_MODEL = os.getenv("GEN_MODEL", "qwen2.5-coder:7b")
+
+mcp = FastMCP(MCP_NAME)
+qdrant = QdrantClient(url=QDRANT_URL)
+
+
+def embed(text: str) -> list[float]:
+    r = requests.post(
+        "http://127.0.0.1:11434/api/embeddings",
+        json={"model": EMBED_MODEL, "prompt": text},
+        timeout=120,
+    )
+    r.raise_for_status()
+    return r.json()["embedding"]
+
+
+def retrieve(query: str, top_k: int = 5):
+    query_vec = embed(query)
+    return qdrant.search(collection_name=COLLECTION, query_vector=query_vec, limit=top_k)
+
+
+def sanitize(text: str, max_len: int = 1800) -> str:
+    return (text or "").strip()[:max_len]
+
+
+@mcp.tool
+def search_docs(query: str, top_k: int = 5) -> str:
+    hits = retrieve(query, top_k=top_k)
+    if not hits:
+        return "No relevant docs found."
+    lines = []
+    for i, hit in enumerate(hits, start=1):
+        src = hit.payload.get("source", "unknown")
+        txt = sanitize(hit.payload.get("text", ""))
+        lines.append(f"[{i}] source={src}\n{txt}")
+    return "\n\n---\n\n".join(lines)
+
+
+@mcp.tool
+def answer_with_citations(query: str, top_k: int = 5) -> str:
+    hits = retrieve(query, top_k=top_k)
+    if not hits:
+        return "Не нашел релевантных документов."
+
+    context = []
+    citations = []
+    for i, hit in enumerate(hits, start=1):
+        src = hit.payload.get("source", "unknown")
+        txt = sanitize(hit.payload.get("text", ""))
+        context.append(f"[{i}] SOURCE: {src}\n{txt}")
+        citations.append(f"[{i}] {src}")
+
+    prompt = (
+        "Ты отвечаешь только на основе контекста.\n"
+        "Если данных недостаточно, явно скажи об этом.\n"
+        "Добавь ссылки на источники в формате [номер].\n\n"
+        f"Вопрос: {query}\n\n"
+        "Контекст:\n"
+        + "\n\n".join(context)
+    )
+
+    r = requests.post(
+        "http://127.0.0.1:11434/api/generate",
+        json={"model": GEN_MODEL, "prompt": prompt, "stream": False},
+        timeout=240,
+    )
+    r.raise_for_status()
+    answer = r.json().get("response", "").strip()
+
+    return f"{answer}\n\nИсточники:\n" + "\n".join(citations)
+
+
+if __name__ == "__main__":
+    mcp.run()