After 2,000 Stars, We Turned the Knowledge Base into a Knowledge Universe

When a knowledge base becomes a knowledge universe.

SAG has passed 2,000 stars on GitHub. Thank you for your support.

GitHub: https://github.com/Zleap-AI/SAG

If this is your first encounter with SAG, here is a one-sentence introduction: SAG is an open-source AI knowledge base application. Add documents to search them, chat with them, trace answers to their sources and explore a body of knowledge through events, entities and relationships.

Today, we are releasing SAG V2.

This update turns the knowledge base into a searchable, observable and navigable “knowledge universe.” We have also rebuilt SAG's core architecture.

01 | Thank You for Every Contribution

Two thousand stars is more than a number.

It represents stars, issues, suggestions and shares, as well as people who have actually run SAG and integrated it into their workflows. Many V2 changes grew out of that feedback.

So our second message is still: thank you for your interest and support.

Now, let us look at what has changed.

02 | A Knowledge Base Is More Than a Search Box

Traditional knowledge bases usually begin with a question: type it in and wait for an answer.

Real knowledge exploration often works differently. You may not yet know what to ask. You want to see relationships between documents, understand the people, concepts and events involved in an issue, and follow a lead from there.

SAG V2 introduces Explore mode, which opens the knowledge base as an interactive 3D knowledge universe.

Search events and entities directly, rotate and zoom the view, or locate a node. Open an event to see its context, then follow its relationships to the next set of knowledge. Return to the original source whenever you need to verify something.

The graph is not decoration. It provides another entry point into knowledge, letting you begin with relationships even before you have a question.

03 | See Relationships in 2D, Enter a Knowledge Planet in 3D

SAG V2 offers three views of the same knowledge: a list, a 2D graph and a 3D graph.

The 2D graph lays out local relationships. Events, entities and their connections are easy to inspect when following a particular line of inquiry.

The 3D graph helps you observe a larger knowledge network. See how all the events and entities relate in three dimensions, like a knowledge planet.

Moving from retrieving knowledge to exploring it is the experience we most wanted this release to deliver.

04 | A New State of the Art in RAG

SAG stands for SQL-Retrieval Augmented Generation. It is an original retrieval architecture designed to replace traditional RAG and GraphRAG, rather than combine them. It turns documents into events and entities, then uses SQL at query time to dynamically connect the local knowledge relevant to a question. There is no global static graph to maintain in advance.

We conducted a consistent evaluation on public multi-hop question-answering datasets, using the same BGE-Large-EN-v1.5 and Qwen3.6-Flash configuration:

• Three datasets: HotpotQA, 2WikiMultiHopQA and MuSiQue.
• Nine metrics across Recall@1, Recall@2 and Recall@5.
• SAG achieved the best result on eight metrics.
• Average Recall@2 and Recall@5 reached 79.30% and 88.18%, respectively, setting a new RAG state of the art.

The full method, experimental setup and results are open source.

Paper: https://arxiv.org/abs/2606.15971

Reproduction code: https://github.com/Zleap-AI/SAG-Benchmark

05 | An Independent Core, Beyond the UI

V2 uses the independent Python package zleap-sag as its core. Importing, event and entity extraction, indexing and retrieval all happen in that package. The web UI is simply one way to access it.

This brings several practical changes:

• Use the complete SAG knowledge base application.
• Connect your own frontend through REST or OpenAPI.
• Integrate existing applications and agents through an OpenAI-compatible interface or MCP.
• Run pip install zleap-sag and use the Python knowledge engine directly in an enterprise pipeline.

The storage layer is also decoupled from the features above it. Locally, start with SQLite and LanceDB without external databases. In production, switch to PostgreSQL, MySQL, OceanBase, pgvector, Elasticsearch or OceanBase Vector. Changing the database does not require rewriting the knowledge workflows.

06 | State-of-the-Art Results Deserve a Better Experience

Our goal is clear: we want SAG to offer the best experience, results and architecture among open-source knowledge base projects.

Answers should be more accurate and traceable to evidence. Knowledge should be visible, understandable and explorable, rather than hidden behind a search box. Core capabilities should be independent, open and directly reusable by other products.

SAG V2 is now available as an open-source update. Explore the knowledge universe, open issues, tell us what you need and help build it with us.

Thank you again to everyone supporting SAG.

See you in the knowledge universe.

Related links:

GitHub: https://github.com/Zleap-AI/SAG

zleap-sag: https://pypi.org/project/zleap-sag/