tooling_graphrag

Sundai hack – testing GraphRAG for llm tooling decision.

Prerequisites

• Python 3.11+
• uv for reproducible installs

---

1 Set up the environment

uv sync

uv creates a local virtual‑env and installs the locked deps from pyproject.toml / requirements.lock.

---

2 Install & initialise GraphRAG

Follow the official quick‑start ➜ https://microsoft.github.io/graphrag/get_started/

---

3 Add your data

Put source files under:

llm_tools_graph/input/

---

4 Build the index

Update settings.yaml, then run:

graphrag index --root ./llm_tools_graph

---

5 Generate prompt for extracting graph

graphrag prompt-tune \
  --root llm_tools_graph \
  --domain "mcp tooling" \
  --selection-method random \
  --limit 1 \
  --k 1 \
  --no-discover-entity-types \
  --output prompts/

Tuned prompts will be saved to prompts/.

---

6 Test tool query

python main.py

---

To Do

• try with smithery tool data
• link to agent
• use neo4j
• try with local models