Skip to content

product.md

Latest commit

 

History

History
53 lines (37 loc) · 2.74 KB

product.md

File metadata and controls

53 lines (37 loc) · 2.74 KB
Raw

GraphRAG product overview

GraphRAG is a self-hosted research workspace for asking grounded questions across private documents. It combines format-aware extraction, hierarchical vector retrieval, bounded knowledge-graph traversal, and inspectable source evidence.

Who it is for

The current release is designed for a researcher, analyst, student, or small-team owner who wants to:

  • upload PDF, DOCX, TXT, and Markdown files;
  • ask questions across one document, a selected set, or the full corpus;
  • inspect the passages and relationships that informed an answer; and
  • keep document storage and retrieval infrastructure under their control.

The included deployment has two configured accounts: an unrestricted owner and a quota-limited shared demo. It does not provide public signup, organizations, collaboration, or a user database.

Product principles

  1. Evidence is part of the answer, not an afterthought.
  2. Model work is selective, configurable, and observable.
  3. Retrieval should be understandable without exposing implementation noise.
  4. Private corpora remain isolated across every storage and retrieval path.
  5. Failure, retry, reset, and deletion behavior are first-class workflows.

What makes it different

GraphRAG uses two complementary retrieval structures:

  • Hierarchical vectors: approximately 200-token child chunks improve search precision, then expand to approximately 1,000-token parent context for generation.
  • Evidence graph: entities and typed assertions are connected to the exact chunks that support them, allowing bounded graph expansion with provenance.

Document extraction is also selective. Machine-readable text stays on deterministic parsers, while scanned PDF pages and meaningful visual regions can be routed through OpenRouter document OCR or a vision-capable model.

Current capabilities

  • Asynchronous document ingestion through Redis and Taskiq.
  • Native and selective VLM extraction for supported formats.
  • Qdrant child-vector search and Neo4j evidence traversal.
  • Streamed answers with explicit [S#] citations.
  • Query-scoped answer-path visualization and a semantic list alternative.
  • Light, dark, and system themes.
  • Owner and demo workspaces with isolated durable data.
  • Session-scoped encrypted OpenRouter BYOK.
  • Docker Compose local hosting and Railway deployment configuration.

Deliberate limits

  • A few hundred documents rather than horizontal multi-tenant scale.
  • A 50 MB and 500-page default owner upload limit.
  • No self-hosted OCR, VLM, embedding, or answer model.
  • No durable cross-device conversation history.
  • No public signup or collaboration in the current phase.

See Architecture, Access model, and Roadmap for the corresponding technical details.