GovernAI Research Atlas — Semantic search

GovernAI Research Atlas

Open-source platform unifying research papers, repositories and governance resources behind a single semantic vector index.

For
GovernAI
Status
Open source
Sources
OpenAlex · GitHub

What it had to solve

AI-governance research is split across incompatible shelves: papers live in academic indexes, implementations live in code hosts, and policy resources live somewhere else again. Keyword search across them fails twice over — the same idea is named differently in a paper and in a repository, and a keyword match cannot tell a central result from a passing mention. What is needed is retrieval by meaning, over everything, with a ranking that knows the difference.

The decisions that shaped it

One embedding space over several sources

Papers from OpenAlex and repositories from GitHub are embedded with Sentence-Transformers into a single vector space. That is what makes the search cross-source rather than federated: a query is compared against everything at once, so a paper and the code implementing it can surface together.

ChromaDB as the retrieval layer

ChromaDB stores and searches the embeddings. Nearest-neighbour retrieval over one index means the system answers by proximity of meaning, so a query phrased in policy language can still reach a result written in engineering language.

A custom relevance score on top

Raw vector similarity is necessary but not sufficient — it will happily return something that mentions the topic once. A custom relevance score re-ranks the retrieved set so that centrality to the query, not just proximity, decides the order.

FastAPI between the index and the interface

A FastAPI service fronts the index and a React client consumes it, which keeps embedding and ranking on the server where they can change without shipping a new front end.

How the pieces sit together

Client
React + TypeScript search interface
API
FastAPI service exposing query, retrieval and ranking
Ranking
Custom relevance score re-ranking the retrieved set
Retrieval
ChromaDB vector search over Sentence-Transformer embeddings
Ingest
OpenAlex papers and GitHub repositories into one index

Built into the product

Cross-source search

Papers, repositories and governance resources in one result set.

Semantic retrieval

Matching on meaning rather than shared keywords.

Custom ranking

A relevance score deciding order after retrieval.

Open source

Published in full, built on behalf of GovernAI.

What it is built out of

Retrieval
ChromaDB, Sentence-Transformers
API
Python, FastAPI
Client
React, TypeScript
Sources
OpenAlex, GitHub

What it looks like

GovernAI Research Atlas — An atlas being built — retrieval and semantic ranking, as five explicit steps.
An atlas being built — retrieval and semantic ranking, as five explicit steps.
GovernAI Research Atlas — The entry point: state an objective, and the atlas is built around it.
The entry point: state an objective, and the atlas is built around it.
GovernAI Research Atlas — Papers ranked by semantic relevance and citation impact, not by date.
Papers ranked by semantic relevance and citation impact, not by date.
GovernAI Research Atlas — The repository side of the index, ranked against the topic rather than by stars.
The repository side of the index, ranked against the topic rather than by stars.
GovernAI Research Atlas — The knowledge graph: papers, repositories, datasets and models around one topic.
The knowledge graph: papers, repositories, datasets and models around one topic.

What it produced

One
Index across papers and repositories
Open source
Published in full, built for GovernAI
Custom
Relevance score over vector retrieval