AtlasStart research

Atlas. Private research,
local-first.

Ask a question, attach a PDF, and get a concise, cited answer.

  • Browser-local fallback
  • No account required
  • No research history

PDFs up to 10 MB

Local research pipeline

PDF in. Cited answer out.

Used by Atlas’s browser-local research path.

  1. 01

    Parse

    PDF.js, page by page.

  2. 02

    Chunk

    Chunks never cross a page.

  3. 03

    Embed

    all-MiniLM-L6-v2, 384 dims.

  4. 04

    Retrieve

    0.8 cosine + 0.2 BM25, top 5.

  5. 05

    Reason

    Five agents, one local model.

  6. 06

    Cite

    Unretrieved citations are dropped.

Answer engine

Atlas tries the hosted OpenAI path when enabled. If it is unavailable, Atlas automatically falls back to local inference in the browser.

Hosted inference
OpenAI · gpt-4.1-mini
Fast answer generation when enabled.
Local fallback
SmolLM2-360M-Instruct · WebGPU / WASM
Runs in the browser when hosted inference is unavailable.

No model selection required. Atlas chooses the available inference path automatically.

Agents

Five roles. One bounded loop.

The multi-role research loop powers the browser-local fallback path.

  1. 1

    Planner

    Up to 4 sub-questions.

  2. 2

    Researcher

    2 validated tool calls per turn.

  3. 3

    Analyst

    Evidence versus inference.

  4. 4

    Critic↺ loop

    Can search again, 2 iterations max.

  5. 5

    Synthesizer

    Answer with page citations.

Under the hood

Small models. Hard limits.

Local inference
SmolLM2-360M-Instruct
q4f16 on WebGPU · q4 on WASM
Context window
8,192
tokens
Embeddings
384-d
all-MiniLM-L6-v2 · q8
PDF limit
10 MB
80 pages · 320,000 chars
Index
400
chunks, in memory
Agent loop
2×
bounded iterations
Hosted inference
gpt-4.1-mini
OpenAI · when enabled

Current limits

  • Text PDFs only, with no OCR for scanned pages.
  • Up to 10 MB and 80 pages per PDF.
  • Hosted answers read only the first part of a long document.
  • Local speed depends on your device, and the first local run downloads the model.
  • Hosted and local answers can differ.

Benchmark

Real runs. Real numbers.

Live tests on the deployed app, October 2026.

Atlas answer screen for “what is the future of ai according to this pdf?” on Impact of AI Until 2050.pdf, 13 pages. Sections: Economic and Job Market Impact; Social, Ethical, and Cultural Implications.Full size ↗

Tap the screenshot to open it at full size.

01 · Live test · Hosted OpenAI path

“what is the future of ai according to this pdf?”

Impact of AI Until 2050.pdf13 pages

Pages analyzed
13
Passages retrieved
4
Citations used
4
Total time
5.44 s
Processing
OpenAI, with the local model as fallback

Sections in the answer

  • Economic and Job Market Impact
  • Social, Ethical, and Cultural Implications
Original status panel: Pages analyzed 13, Passages retrieved 4, Citations used 4, Total time 5.44 s, Processing OpenAI, with the local model as fallback.
Original status panel from the same run.

Ask a question. Keep your data.

No account, research history, or stored documents. Sessions live in memory and clear on refresh or close. Local inference keeps research in the browser. With hosted inference enabled, your question, file name, and extracted text are processed through OpenAI on the server.

Start research