litegraf 0.1.0

Lightweight knowledge graph ingestion and enrichment pipeline

LiteGraf

Lightweight knowledge graph ingestion and query pipeline. Insert text or documents, extract entities and relationships with an LLM, store them in a graph database, and query with natural language.

from pipeline.litegraf import LiteGraf

kg = LiteGraf()
kg.insert("TP53 is associated with multiple cancers including breast and lung cancer.")
result = kg.query("What cancers are associated with TP53?")
print(result.answer)

Features

• Single entry point — LiteGraf dataclass with sensible defaults, override only what you need
• Pluggable backends — Neo4j, Memgraph, Ollama, Cloudflare Workers AI, AWS Bedrock
• Sync and async — insert() / ainsert(), query() / aquery()
• Content deduplication — hash-based, idempotent inserts
• LLM response caching — disk-based, avoids redundant API calls
• Rate limiting — async concurrency limiter for LLM providers
• PDF and document ingestion — via MarkItDown + PyMuPDF
• Benchmarking suite — compare extraction quality across LLM providers
• Enrichment pipeline — entity resolution, ontology integration, evidence scoring

Install

Requires Python 3.11+.

pip install litegraf

With optional backends:

pip install litegraf[neo4j]       # Neo4j graph store
pip install litegraf[bedrock]     # AWS Bedrock LLM
pip install litegraf[all]         # Everything

Or from source with uv:

git clone https://github.com/graffold/litegraf.git
cd litegraf
uv sync --all-extras

Quick Start

Default setup (Ollama + Neo4j)

Start Ollama and Neo4j locally, then:

from pipeline.litegraf import LiteGraf

kg = LiteGraf()  # connects to localhost defaults

# Insert text
kg.insert("BRCA1 interacts with RAD51 in DNA repair pathways.")

# Insert a PDF
kg.insert(open("paper.pdf", "rb").read())

# Query
result = kg.query("What proteins interact with BRCA1?")
print(result.answer)
print(result.context)  # retrieved graph context

Cloudflare Workers AI (free tier)

kg = LiteGraf(
    llm="cloudflare",
    llm_model="@cf/meta/llama-3.1-8b-instruct-fp8",
)

Memgraph backend

kg = LiteGraf(
    graph_store="memgraph",
    graph_uri="bolt://localhost:7687",
    graph_user="",
    graph_password="",
)

Async usage

import asyncio
from pipeline.litegraf import LiteGraf

async def main():
    kg = LiteGraf()
    await kg.ainsert("TP53 suppresses tumor growth.")
    result = await kg.aquery("What does TP53 do?")
    print(result.answer)

asyncio.run(main())

Query modes

# Full pipeline: retrieve context → LLM synthesis
result = kg.query("What cancers involve TP53?")

# Context only (bring your own LLM prompt)
result = kg.query("TP53", mode="only_context")
for chunk in result.context:
    print(chunk.text, chunk.score)

Configuration

All parameters can be set via the LiteGraf constructor:

Parameter	Default	Description
graph_store	"neo4j"	Graph backend: "neo4j", "memgraph", or instance
graph_uri	"bolt://localhost:7687"	Bolt connection URI
graph_user	"neo4j"	Graph database username
graph_password	""	Graph database password
llm	"ollama"	LLM provider: "ollama", "cloudflare", "bedrock"
llm_model	"llama3"	Model name/ID
embedding	"local"	Embedding provider (local sentence-transformers)
chunk_token_size	512	Tokens per chunk
enable_cache	True	Cache LLM responses to disk
enable_dedup	True	Skip duplicate content on insert

Benchmarks

Compare extraction quality across LLM providers on biomedical datasets:

python -m pipeline.benchmarks

Results are published to docs/ for GitHub Pages viewing.

Development

uv sync --all-extras --group dev
uv run pytest
uv run ruff check src/

License

AGPL-3.0