Graph-aware retrieval, now in general availability

Map your knowledge
like the stars.

Startrag turns flat document chunks into a navigable constellation of meaning. Graph-aware retrieval that finds the right context, boosts accuracy, and cuts hallucinations.

$ pip install startrag

Trusted by AI teams building on

NorthwindQuantaHelios AIMeridianVela Labs
See the difference

Drag to compare retrieval strategies

Traditional RAG retrieves isolated chunks by raw vector distance. Startrag walks a knowledge graph of clustered stars — so answers stay grounded, complete, and connected.

Traditional RAG
Starfield RAG

Answer accuracy

86.1%

Hallucination rate

7.1%

Context recall

82%

The star map

Retrieval that understands how knowledge connects

Startrag rebuilds your knowledge base as a living star chart, so your models retrieve meaning — not just nearby vectors.

Chunks become stars

Every document slice is embedded and placed as a star, sized by importance and colored by topic — never a flat list again.

Vector clusters as constellations

Semantically related stars are linked into constellations, capturing the relationships traditional cosine distance ignores.

Graph-walk retrieval

Queries traverse the star map, pulling complete neighborhoods of context instead of disconnected fragments.

Grounded by design

Every answer carries citations back to its source stars, so you can audit exactly where the model looked.

Sub-100ms recall

A purpose-built graph index returns rich, connected context in milliseconds, even across millions of stars.

Drop-in compatible

Swap your existing vector store with two lines of code. Works with any model, framework, or embedding.

How it works

From raw documents to a navigable galaxy

01

Ingest & slice

Point Startrag at your docs, wikis, and databases. We chunk and embed every source into stars.

02

Cluster & link

Stars are clustered into constellations and connected by semantic edges to form your knowledge graph.

03

Traverse & ground

At query time, Startrag walks the graph to assemble complete, cited context for your model.

04

Answer with confidence

Your LLM responds with grounded, traceable answers — and measurably fewer hallucinations.

98.4%

Answer accuracy

−91%

Fewer hallucinations

<100ms

Median retrieval

12M+

Stars indexed / cluster