scripts/ollama_tool_runner.py hinzugefügt

scripts/ollama_tool_runner.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+scripts/ollama_tool_runner.py — Minimaler Tool-Caller für Ollama + mindnet
+
+Zweck:
+    Führt den Tool-Call-Loop für mindnet_query / mindnet_subgraph aus.
+    Nutzt /tools/ollama (Schema) und deine FastAPI (/query, /graph).
+ENV:
+    OLLAMA=http://127.0.0.1:11434
+    API_BASE=http://127.0.0.1:8000
+    MODEL=llama3.1
+Nutzung:
+    python3 scripts/ollama_tool_runner.py "Frage an den Agenten"
+Version:
+    0.1.0  (Erstanlage)
+Stand:
+    2025-10-07
+"""
+
+from __future__ import annotations
+import os, sys, json, requests
+
+OLLAMA = os.getenv("OLLAMA", "http://127.0.0.1:11434")
+API_BASE = os.getenv("API_BASE", "http://127.0.0.1:8000")
+MODEL = os.getenv("MODEL", "llama3.1")
+
+def load_tools():
+    r = requests.get(f"{API_BASE}/tools/ollama", timeout=30)
+    r.raise_for_status()
+    return r.json()["tools"]
+
+def call_model(messages, tools):
+    payload = {
+        "model": MODEL,
+        "messages": messages,
+        "tools": tools,
+        "stream": False,
+        "tool_choice": "auto",
+    }
+    r = requests.post(f"{OLLAMA}/api/chat", json=payload, timeout=180)
+    r.raise_for_status()
+    return r.json()
+
+def execute_tool(name: str, arguments: dict):
+    if name == "mindnet_query":
+        # mappe expand_depth/edge_types auf /query-Body
+        body = {
+            "mode": "hybrid",
+            "top_k": int(arguments.get("top_k", 10)),
+            "expand": {
+                "depth": int(arguments.get("expand_depth", 1)),
+                "edge_types": arguments.get("edge_types") or ["references","belongs_to","prev","next"],
+            },
+            "filters": arguments.get("filters"),
+        }
+        if "query_vector" in arguments:
+            body["query_vector"] = arguments["query_vector"]
+        if "query" in arguments:
+            body["query"] = arguments["query"]
+        r = requests.post(f"{API_BASE}/query", json=body, timeout=120)
+        r.raise_for_status()
+        return r.json()
+
+    if name == "mindnet_subgraph":
+        params = {
+            "depth": int(arguments.get("depth", 1)),
+        }
+        edge_types = arguments.get("edge_types")
+        if edge_types:
+            # mehrfacher Query-Param 'edge_types' wird von requests automatisch gebaut
+            params["edge_types"] = edge_types
+        note_id = arguments["note_id"]
+        r = requests.get(f"{API_BASE}/graph/{note_id}", params=params, timeout=60)
+        r.raise_for_status()
+        return r.json()
+
+    raise ValueError(f"Unbekanntes Tool: {name}")
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: python3 scripts/ollama_tool_runner.py \"Deine Frage\"")
+        return 2
+
+    user_prompt = sys.argv[1]
+    tools = load_tools()
+    messages = [{"role": "user", "content": user_prompt}]
+
+    while True:
+        resp = call_model(messages, tools)
+        msg = resp.get("message") or {}
+        tool_calls = msg.get("tool_calls") or []
+
+        if not tool_calls:
+            # finale Antwort
+            print("\n=== Antwort ===\n")
+            print(msg.get("content", "").strip())
+            return 0
+
+        # Führe Toolcalls nacheinander aus
+        for tc in tool_calls:
+            name = tc["function"]["name"]
+            args = tc["function"].get("arguments") or {}
+            if isinstance(args, str):
+                try: args = json.loads(args)
+                except Exception: args = {"raw": args}
+
+            print(f"[tool] {name}({args})")
+            result = execute_tool(name, args)
+
+            # Übergib Toolresultat ans Modell zurück
+            messages.append({"role": "tool", "content": json.dumps(result), "name": name})
+
+        # füttere auch die originale Antwort (falls enthalten) in den Verlauf
+        if msg.get("content"):
+            messages.append({"role": "assistant", "content": msg["content"]})
+
+if __name__ == "__main__":
+    sys.exit(main())