Lumen-Labs/brainapi2

What it solves

BrainAPI transforms unstructured raw text into a structured, queryable knowledge graph automatically. It solves the problem of "similarity guesses" in traditional RAG by providing answers backed by traceable, reasoned paths through a graph, allowing for multi-hop reasoning and a clear provenance trail for every fact.

How it works

The system uses a "swarm" of specialized AI agents to process text:

• Scout: Extracts all entities (people, things, events).
• Observations Agent: Writes factual notes based on the text and prior context.
• Architect: Creates event-centric relationships (subject $\rightarrow$ event $\rightarrow$ object) where actions are treated as central hubs.
• Janitor: Validates the relationships, checks for duplicates, and loops back to the Architect for corrections until the data is clean.

Knowledge is stored in an append-only graph, ensuring that the history of how facts evolved is preserved. It offers two pipeline modes: accurate (full swarm validation) and lightweight (faster, basic extraction).

Who it’s for

• AI Agent Developers: To provide persistent, structured memory with provenance.
• Data Analysts: To extract business intelligence from qualitative data like support conversations.
• Researchers/Investigators: To map connections between people, events, and documents for compliance or investigation.
• Recommendation System Engineers: To build recommendation engines based on behavioral paths rather than simple co-occurrence.

Highlights

• Event-Centric Modeling: Captures who did what, to whom, and when, rather than simple flat edges.
• Traceable Answers: Provides the full graph path (triples) used to derive an answer.
• Agentic Pipeline: Uses a multi-agent swarm to ensure data quality and deduplication.
• Extensible via Plugins: Allows users to override prompts, add REST routes, or register MCP tools for specific domains (e.g., CRM, legal).
• MCP Support: Integrates with Claude Desktop via a Model Context Protocol server.