SemAtlas

Ask a question and watch Wikipedia answer it. Live at sematlas.com/q.

SemAtlas searches Wikipedia, follows the links most likely to lead to an answer, and writes a short Wikipedia-style article from what it found, citing every article it used. The search happens in the open: you watch the graph of articles grow, see why each one was followed, and decide for yourself which articles the answer should be built from.

Most AI answer engines hand you a conclusion and a list of links. SemAtlas is built around the opposite idea: the path to an answer is worth seeing, and the reader, not the machine, should get the last word on what counts as evidence.

What you can do

• Ask a question. The graph grows as articles are chosen. A node's colour says which starting article it grew from; its brightness says how strongly it answers the question. Dashed boxes are links that were ranked but not followed.
• Choose the evidence. Click any article, followed or not, to add it to the answer's sources or take it out, then rewrite the answer.
• Ask follow-ups. The same graph keeps growing, and each answer builds on the last.
• Check everything. Every claim cites its source, every ranking decision is listed with its score, and the exact models and the cost of each answer are shown.
• Explore topics without a question on the original page at /, which draws the link graph around any Wikipedia article.

How it works

question
  → Wikipedia search (20 results)
  → Jev picks the starting articles
  → best-first growth, a few rounds: the most promising article's lead links,
    with their one-line descriptions, are ranked against the question; the
    best are followed, the next few stay in the graph as candidates
  → every article's intro is scored as evidence; the most promising are read
    in full and every paragraph is scored
  → a chat model writes the answer from the strongest intros and paragraphs,
    citing them as numbered references

Ranking is done by Jev, a "decision model" that returns calibrated probabilities instead of text: it is fast (well under a second for a hundred judgements) and costs $0.042 per million input tokens. Writing is done by a general chat model, openai/gpt-5.4-mini by default. A question costs about 1.5 cents and takes about 12 seconds. The code is in backend/explore.py (the exploration), backend/ask.py (scoring and writing) and static/js/q.js (the page).

Run it yourself

Python 3.11 or newer.

git clone https://github.com/josephbakarji/sematlas
cd sematlas
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add an OpenRouter key: https://openrouter.ai/keys
hypercorn main:app --bind 127.0.0.1:5000

Then open http://127.0.0.1:5000/q.

One OpenRouter key covers both models. You can instead point the writer at OpenAI, Anthropic, or a local OpenAI-compatible server such as Ollama (set SEMATLAS_LLM=openai and OPENAI_BASE_URL); see .env.example. Without Jev, relevance is scored by the writing model instead, which works but is slower, and the graph does not grow past its starting articles.

Running a public copy

Every question costs money, so a public deployment needs to decide who pays. SemAtlas gives each visitor a few free questions a day, paid by the operator's key inside a daily dollar budget (SEMATLAS_FREE_PER_DAY, SEMATLAS_FREE_BUDGET_USD). Beyond that, visitors connect their own OpenRouter account with one click (OAuth with PKCE) or paste a key. Their key stays in their browser, is sent only with their own requests, and is never stored or logged by the server.

Please set SEMATLAS_CONTACT to a way to reach you: Wikimedia asks every API client to identify its operator.

The repository deploys as is to Render or any host that runs a Procfile or the included Dockerfile.

Privacy

The server logs question text, for debugging. It keeps anonymous counts of what people do on the page (event names such as "added a source" or "opened full view", with no identifiers) at /q/stats. Keys supplied by visitors are never logged or stored.

Licence

Code: MIT, see LICENSE. Wikipedia text is by Wikipedia contributors under CC BY-SA 4.0, and answers written from it are shared under the same licence.