---

name: nano-graphrag description: "Use nano-graphrag for lightweight GraphRAG indexing, querying, provider customization, storage backend selection, prompt/entity extraction tuning, and troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: MIT

nano-graphrag

Use this repo skill when a task involves the nano-graphrag Python package, lightweight GraphRAG/RAG indexing, graph-backed retrieval, local/global/naive query modes, custom LLM or embedding providers, storage backend swaps, entity-extraction troubleshooting, or GraphML knowledge graph inspection.

The root stays a router. Open the focused sub-skill for the workflow you need, then use its bundled references and scripts instead of reopening the source repository.

Quick environment check

Install the package in the active project environment with one of:

pip install nano-graphrag
# or, for a checkout you are actively editing:
pip install -e .

Then verify importability:

from nano_graphrag import GraphRAG, QueryParam
print(QueryParam(mode="global"))

If import fails with ModuleNotFoundError: No module named 'transformers', install transformers; this repository version imports transformers.AutoTokenizer at module import time even though the package metadata may not list it explicitly.

Run scripts/check_nano_graphrag_env.py when you need a safe no-network import/API/storage sanity check before following deeper recipes.

Route by task

User task	Read first
Build a GraphRAG, insert text, query global/local/naive modes, use async methods, reuse a working_dir, or validate custom chunking.	sub-skills/core-graphrag-workflows/SKILL.md
Replace OpenAI with Azure OpenAI, Amazon Bedrock, DeepSeek/OpenAI-compatible APIs, Ollama, local embeddings, or debug provider kwargs/credentials.	sub-skills/provider-and-model-integrations/SKILL.md
Choose JSON/NanoVectorDB/HNSW/NetworkX/Neo4j storage, adapt FAISS/Milvus/Qdrant-style vector stores, or visualize GraphML.	sub-skills/storage-backends/SKILL.md
Change prompts, repair malformed JSON, customize entity extraction/DSPy, or recover from zero entities / empty graph / Leiden failures.	sub-skills/customization-and-troubleshooting/SKILL.md

Read these root references

• Read references/package-overview.md for package purpose, public API surface, default components, install caveats, and how the sub-skills fit together.
• Read references/troubleshooting.md for cross-cutting install/import, provider, storage, empty-graph, and optional dependency triage before choosing a deeper troubleshooting page.
• Read references/repo-provenance.md before deciding whether this skill is current for a checkout or whether refresh-repo-skill is needed.
• references/repo-routing-metadata.json is structured metadata consumed by the repo-skills router import process; do not edit it casually.

High-confidence operating rules

1. Use a stable working_dir for persistent state. The default local stack writes JSON KV stores, NanoVectorDB JSON files, and NetworkX GraphML under that directory.
2. Do not run hosted-provider examples until credentials, network policy, model IDs, and rate limits are explicit. For no-network validation, use fake LLM/embedding helpers from the core sub-skill.
3. Enable enable_naive_rag=True before insertion if you will later query QueryParam(mode="naive").
4. Treat Neo4j, Ollama, Bedrock, FAISS, Milvus, Qdrant, and downloaded local embedding models as optional integrations with their own services, packages, or credentials.
5. If insertion produces no entities/relations, debug provider output format and context length before replacing graph storage.
6. Keep any generated application scripts independent of this source checkout; copy or adapt needed patterns into the application or use the bundled scripts here.