🧠 I built a fully local Graph RAG CLI for Obsidian vaults using LlamaIndex + Ollama

[benmaster82]

Hey LlamaIndex community 👋

I wanted to share a project I've been working on that's built almost entirely on top of LlamaIndex's graph capabilities - a local-first, privacy-respecting knowledge graph you can query in natural language from your terminal.

What is it?

Kwipu is a Graph RAG system that turns your Markdown notes - especially Obsidian vaults - into a queryable knowledge graph, with zero cloud dependency.

🔗 GitHub: https://github.com/benmaster82/Kwipu

How LlamaIndex powers it

The core of the project relies heavily on:

• PropertyGraphIndex - to build and persist the knowledge graph from extracted triples
• SimpleLLMPathExtractor - for entity-relation extraction from free text via Ollama
• ImplicitPathExtractor - for implicit relationships between nodes
• LLMSynonymRetriever + VectorContextRetriever - combined for hybrid retrieval

On top of that I added two custom retrievers:

• A BM25ChunkRetriever with multilingual tokenizer support (EN, IT, FR, DE, ES, PT)
• A TemporalMetadataRetriever for date/tag-based lookup

The hybrid retrieval strategy

The system fuses 4 retrieval signals at query time:

Query
  ├── LLMSynonymRetriever     (query expansion via LLM)
  ├── VectorContextRetriever  (semantic similarity, nomic-embed-text)
  ├── BM25ChunkRetriever      (keyword scoring)
  └── TemporalMetadataRetriever (dates, tags, events)

Results are merged and passed to a tightly constrained generation prompt that enforces source citation - no hallucinations about content not in the graph.

Obsidian-native pre-processing

Before the LLM extractors even run, a custom pre-processor parses:

• [[wikilinks]] → structural triples
• YAML frontmatter → metadata nodes

This means even small/cheap models produce a rich graph because a lot of the structure is already explicit in the notes.

Stack

Component	Tool
Graph index	LlamaIndex PropertyGraphIndex
LLM backend	Ollama (local)
Embeddings	nomic-embed-text via Ollama
File watching	Watchdog + MD5 hash diffing
Concurrency	Custom ReadWriteLock
UI	Rich (terminal)

Questions / things I'd love feedback on

1. Incremental updates - right now file modifications trigger a full rebuild. I'm using insert_document for new files, but modifying an existing node cleanly is trickier. Has anyone found a good pattern for partial graph invalidation with PropertyGraphIndex?

2. Retriever fusion - I'm doing a naive merge of the 4 retrievers. Is there a built-in re-ranking or score normalization utility in LlamaIndex I should be using instead?

3. Graph persistence - I'm persisting with StorageContext.from_defaults(persist_dir=...). Any gotchas with large graphs (500+ notes)?

Would love to hear thoughts from anyone who's built something similar or has experience with LlamaIndex's graph layer at scale.

Thanks for building such a solid framework 🙏

Built with: LlamaIndex · Ollama · Python · Obsidian