SQL-RAG · EVENT-ENTITY RETRIEVAL

SAG

Searchable. Connected. Traceable.

Turn scattered documents into structured knowledge and source-grounded Agent answers. SAG is a local-first, open-source knowledge base built around an original retrieval architecture instead of stitching traditional RAG and GraphRAG together.

Read the docs GitHub

Local-first MIT licensed Python 3.11+ v1.2.2

Product overview

Two perspectives. One traceable knowledge base.

Use the workspace for ingestion, retrieval, and citation review, then enter Explore mode to unfold the entire knowledge base into a universe you can move through freely.

Workspace

Why SAG

Keep complete semantics in Events. Build relationships at query time.

SAG neither maintains two retrieval systems nor precomputes global hyperedges. For the current question, SQL connects events through shared entities to create a focused local candidate space.

ChunkSource-evidence boundary

→

EventComplete semantic unit

↔

EntitiesIndex and expansion points

→

Dynamic hyperedgeBuilt locally at query time

A verifiable knowledge path

Ingest once. Preserve evidence through retrieval and reuse.

Import

Normalize files and web pages to Markdown

Index

Build Chunk, Event, and Entity representations

Retrieve

Use semantic recall and query-time hyperedges

Trace

Return every result to its source evidence

Reuse

Connect Agents, APIs, MCP, and Python

System boundary

One engine. Three entry points.

Use the complete workspace, build a custom interface on the self-hosted API, or embed zleap-sag directly in a Python service. Every path shares the same knowledge engine.

SAG Desktop / WebREST / OpenAI / MCPzleap-sag

SAG knowledge engineOne data model for indexing, retrieval, and citations

SAG Desktop / WebA ready-to-use knowledge workspace REST / OpenAI / MCPSelf-hosted integration boundaries zleap-sagAn embeddable Python engine

Complete product

Search, converse, and explore the same knowledge.

Retrieval and tracingFast / precise modes

Inspect retrieval before an Agent uses it, and open the source behind every result.

Cited conversationsStreamed Agent answers

Select knowledge scope, observe tool activity, and keep important claims attached to clickable citations.

Knowledge UniverseEvent-Entity exploration

Discover relationships across events, entities, and timelines, then return from any node to source evidence.

Public benchmarks

Best in 8 of 9 Recall@1/2/5 metrics.

With the same BGE-Large-EN-v1.5 Embedding and Qwen3.6-Flash LLM configuration, SAG reports average Recall@2 of 79.30% and Recall@5 of 88.18% across three multi-hop QA datasets.Read the paper and full results

Recall@279.30%HippoRAG 2 · 68.14%

Recall@588.18%HippoRAG 2 · 83.28%

Best metrics8 / 93 public datasets

Start with SAG

Turn your first document into verifiable knowledge.

Desktop runs out of the box, Docker starts in three lines, and the Python package embeds directly.

Choose an installation path