Demo edition · a research pipeline, streamed

Meridian

Ask a research question and watch the pipeline work on it: sources found, ranked, critiqued and cited.

Python · FastAPI · asyncio · Server-Sent Events Wikipedia & arXiv

This project at a glance

The problem
One search and one chunk rarely answer a real question, and an LLM answer rarely shows where it came from.
What I built
A five-agent research API: plan sub-questions, search Wikipedia and arXiv concurrently, rank passages with embeddings, critique, write a cited report. Every stage streams live over SSE.
Evidence
Every report gets a grounding check: each sentence is matched against the passages it cites, and unsupported ones are listed. 53 tests, CI on every push.
Methods
Python, FastAPI, asyncio fan-out, Server-Sent Events, embeddings, SQLite sessions, Groq (optional).

1Press Research. 2Watch the lab notebook fill in, stage by stage. 3Read the fact check under the report.

Or try
    Report

    The report will be set here once the pipeline finishes. A run takes a few seconds; the lab notebook records each stage as it happens.

    §1How it works

    Five stages in an asyncio pipeline. The sub-questions are searched concurrently with asyncio.gather, so one slow or failing source cannot stop the run; each stage pushes events to a queue that the HTTP response streams out.

    Question to planner, then three concurrent researchers, then extractor, critic and synthesizer, then a cited report; every stage streams events. Question Planner LLM: sub-questions Researcher Researcher Researcher Extractor embed + cosine rank Critic LLM: gaps, conflicts Synthesizer LLM: cited report Report citations checked Wikipedia + arXiv (+ DuckDuckGo if few) every stage emits events → streamed as SSE

    §2Appendix: stream it yourself

    The page reads the same endpoint. Public use is limited to a few questions an hour per visitor.