Lars
d797957abc
All checks were successful
Deploy mindnet to llm-node / deploy (push) Successful in 3s
120 lines
3.8 KiB
Python
120 lines
3.8 KiB
Python
#!/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())
|