---
name: query-hf-graphrag
description: Query the CID-keyed ipfs_kenya_laws_ir GraphRAG release on Hugging Face with bounded BM25, centroid-routed vectors, adjacency neighbors, and graph walks. Use when an agent must retrieve lexical, semantic, or graph context without downloading the full corpus.
---

# Query ipfs_kenya_laws_ir on Hugging Face

Use the bundled search script. It fetches only the manifest, compact locators,
matching posting/vector/adjacency shards, and corpus rows for final hits.

Repository: `justicedao/ipfs_kenya_laws_ir`. Primary key is `entry_cid`. Integer `document_index`
is a compact pointer, not an identity.

Research snapshot. Similarity edges cannot establish legal authority.

## Choose retrieval

- `bm25` for citations, docket numbers, quoted terms, and exact names.
- `vector` for paraphrases and topical similarity.
- `neighbors` / `walk` after you have a `node_cid`.
- Merge BM25 and vector hits on `entry_cid`. Similarity edges (`BM25_NEIGHBOR_OF`, `EMBEDDING_NEIGHBOR_OF`, `contains_term`) are retrieval proposals only.

## BM25

```bash
python scripts/query_hf_graphrag.py \
  --repo-id justicedao/ipfs_kenya_laws_ir \
  --revision <pinned-commit> \
  bm25 "clean air act section 111" --top-k 10
```

Local checkout:

```bash
python scripts/query_hf_graphrag.py --local-root . bm25 "clean air act section 111"
```

Install `pyarrow` and `huggingface_hub`. Set `HF_TOKEN` for private datasets.
Always pin `--revision` for remote queries; never use `main`.

## Neighbors and walks

```bash
python scripts/query_hf_graphrag.py --local-root . neighbors <node-cid> --direction both --limit 25
python scripts/query_hf_graphrag.py --local-root . walk <node-cid> --max-depth 2 --max-nodes 100
```

Inspect `fetch_trace`. A budget stop is a partial walk, not proof that no
further path exists. Treat hits as `context_only`.

Read [references/schema.md](references/schema.md) for locator and posting
contracts.