Open Source · Self-Hosted · Enterprise Ready

Document Intelligence.
Engineered for Precision.

hRAG fuses semantic vector search and BM25 keyword precision into a unified hybrid retrieval engine — private, stateless, and deployable anywhere from a laptop to a multi-node cluster.

Ingest
Extract
Embed
BM25+Vec
RRF Fuse
Re-rank
Answer

Hybrid Retrieval Pipeline — From raw document to AI-grounded answer

The Problem

Why neither pure approach is enough

Pure Vector Search

Great at finding semantically similar content but misses exact matches. Product codes, names, dates, and precise terminology vanish in the embedding space.

  • Misses exact keyword matches
  • Struggles with rare terms
  • Opaque ranking logic
hRAG

Hybrid Retrieval

Reciprocal Rank Fusion merges vector and keyword rankings. A cross-encoder re-ranker then re-scores the merged list for maximum relevance.

  • Exact + semantic coverage
  • Cross-encoder re-ranking
  • Private by default, ACL-enforced

Pure Keyword (BM25)

Precise but brittle. Misses paraphrases, synonyms, and conceptual questions. "What are our data retention obligations?" returns nothing without those exact words.

  • No semantic understanding
  • Synonyms / paraphrase blind
  • Fails conceptual queries

Platform Capabilities

Everything you need. Nothing you don't.

Core Engine

Hybrid Search Engine

Reciprocal Rank Fusion merges LanceDB vector scores with BM25 keyword precision. A cross-encoder re-ranker delivers the final cut.

Ingestion

Intelligent Ingestion

PDF type detection (Native / Scanned / Sandwich), WASM OCR, async worker threads. Zero UI blocking on large documents.

UX

Streaming AI Chat

Server-Sent Events stream answers token-by-token with deep-linked source citations back to the exact document page.

Security

Iron-Clad Multi-Tenancy

Owner-anchored ACL on every vector query. Defense-in-Depth post-filters relational permissions before any result reaches a client.

Intelligence

Auto-Classification

High-Water Mark logic locks documents to the highest severity between user input and AI assessment. Local, Ollama, or Cloud.

Architecture

Stateless Architecture

Compute nodes hold zero state. Garage S3 stores objects and vectors; Litestream replicates SQLite. Deploy anywhere, scale freely.

Deployment Options

From laptop to enterprise cluster.

One codebase. Three deployment tiers. No vendor lock-in.

Free

Local

Zero infrastructure. Run on any machine with Node.js. SQLite + local LanceDB. Perfect for personal use or evaluation.

Node.jsSQLiteLanceDB (local)
Get Started Free →
Recommended

Self-Hosted

Production-grade on a single VPS. Garage S3 for object storage, Litestream for SQLite replication, LanceDB S3-native vectors.

Garage S3LitestreamLanceDB (S3)
View Deployment Guide →
Enterprise

Enterprise

Horizontal multi-node cluster with primary election, PostgreSQL + pgvector, BullMQ distributed ingestion, and full HA.

PostgreSQLBullMQpgvector / S3
Read Architecture Docs →

Iron-Clad Security

Security isn't a feature.
It's the foundation.

Every retrieval is ACL-enforced at the vector layer, then cross-verified at the relational layer before any result reaches the client.

Full Security Architecture →
AES-GCM secrets at rest, derived from master passphrase
Pre-signed S3 URLs with 60-second TTL — no long-lived tokens
Vector ACL hard-filter + relational post-filter (Defense-in-Depth)
JWT-anchored WHERE on every relational query
Traceable delegation — every share/revoke captured in audit log
High-Water Mark classification locks cannot be downgraded

Take control of your
document intelligence.

Self-host in minutes. Full source available. No SaaS dependency, no data leaving your infrastructure.