Platform

Platform Features

A complete document intelligence platform. From ingestion to retrieval to streaming AI answers — every component built for production.

Search & Retrieval

Hybrid Search (RRF)

Core

Vector + BM25 ranked lists merged with Reciprocal Rank Fusion. Neither approach alone — both working together.

Cross-Encoder Re-ranking

Precision

Top-N candidates jointly scored against the raw query by a cross-encoder model. The most accurate, most expensive step applied last.

LanceDB Vector Store

Storage

S3-native columnar vector store with sub-100ms ANN retrieval. Tiered abstraction separates the embedding provider from the store.

Full-Text BM25 Index

Index

LanceDB built-in FTS keeps keyword and vector search in the same store. Zero extra infrastructure.

Document Ingestion

PDF Type Detection

Intelligence

Structural heuristics detect Native, Scanned, and Sandwich PDFs automatically — choosing the optimal extraction path per document.

WASM OCR (Tesseract)

OCR

In-process OCR via tesseract.js. Page-level rasterization pipeline with progressive status reporting. Force OCR override available.

Office Document Extraction

Formats

DOCX, PPTX, XLSX, ODT and legacy formats (.doc, .xls). Embedded images extracted and OCR-processed separately.

Async Worker Threads

Performance

Ingestion runs in Node.js Worker Threads (or BullMQ in Enterprise). Zero UI blocking. Auto-classification included in the async pass.

AI & Chat

Streaming SSE Chat

UX

Answers stream token-by-token to the browser via Server-Sent Events. No waiting for the full response before reading begins.

Deep-Linked Source Citations

Transparency

Every answer links back to the exact document, page, and passage used. Click to jump directly to the source.

Multi-Model Support

Flexibility

Connect Local (WASM), Ollama, or Cloud LLMs. Embedding provider is independently configurable from the chat LLM.

Auto-Classification Engine

Compliance

High-Water Mark: documents are locked to the highest severity between user-declared and AI-assessed classification. Irreversible downgrade protection.

Security & Access Control

Private-by-Default ACL

Iron-Clad

Every vector query is anchored on owner_id. Creators always have access even if no explicit sharing exists. Hard-filter at the engine level.

Defense-in-Depth

Zero-Trust

Vector ACL filter at retrieval + relational permission cross-check before delivery. Two independent layers, both must pass.

Pre-Signed S3 Downloads

Downloads

Document downloads use authenticated pre-signed S3 URLs with a strict 60-second TTL. No long-lived access tokens.

AES-GCM Secret Storage

Secrets

All secrets encrypted at rest in S3 using AES-GCM, derived from the operator's master passphrase. Decrypted in-memory on boot only.

Administration

Fleet Health Dashboard

Ops

All nodes report heartbeats every 60 seconds to the shared registry. Ops team sees CPU, memory, uptime, primary/replica status at a glance.

Audit Log

Compliance

Every share, revoke, classification change, and admin action captured as a high-priority audit event. Immutable append-only log.

System Settings Registry

Stateless

All system-wide configuration lives in the relational DB (system_settings table). Nodes cache with a short TTL — no config files on disk.

Multi-Tenant Groups

Tenancy

Users belong to one or more groups with group-scoped roles. Cross-group users are supported. Privacy scoping enforced at every layer.

See it for yourself.

Deploy locally in under 5 minutes. No cloud account required.