From source material to useful answers.
Bring a webpage or text document. Inspect the passages, choose your model, and ask a question. Every generated answer comes from your selected provider, through a server-side proxy using your own key.
Your provider, your key.
NOT VERIFIEDYour key is sent to this site’s server only when you generate, then forwarded to the selected provider over HTTPS. It is never saved by this app in cookies, browser storage, files, or a database. Reload or clear the session to forget it.
Model presets are editable starting points; availability and billing depend on your account. View provider models ↗
Source collection
0 / 5Start with something worth asking about.
Add your own sources, or explore retrieval with an authored educational example.
Grounded workspace
NO GENERATED DEMO ANSWERSEvidence before answers.
Inspect what will be retrieved. Add your provider key when you’re ready for a real summary or answer.
NO KEY NEEDED FOR SOURCE INSPECTIONRetrieval preview
0 / 0 PASSAGESSummary preview: first passages, rotating across sources. Up to six 1,200-character passages, with 120-character overlap. Semantic/vector search is not connected.
Add a source to see its passages here.
A small research pipeline you can inspect.
Content, without executing a page.
The server fetches public HTTPS HTML with a 1 MB limit, bounded redirects, a 15-second deadline, and DNS-pinned public addresses. It strips navigation and scripts, then extracts readable blocks. JavaScript-only pages, authentication, and paywalls are unsupported.
Real lexical retrieval.
Sources are chunked with overlap. Question terms rank passages using capped term frequency and inverse chunk frequency. Summaries rotate through sources. Only six passages are sent to the model; this can miss relevant context. Embeddings, a vector database, and PDF/OCR are not connected.
A bounded provider request.
OpenAI, Anthropic, and Gemini adapters use fixed HTTPS endpoints, no redirects, a 35-second deadline, an output-token limit, and no automatic retries. Provider keys travel in headers, never URLs. Adapters are covered by mocked contract tests; successful live calls require an available model and your valid key.
Citations are a review trail.
The server rejects unknown passage IDs and copies cited excerpts from the supplied sources. This verifies source membership, not whether an answer is true. Source content is untrusted; models can still follow malicious instructions. No tools or privileged actions are exposed to the model.
Sources, answers, and keys are held in tab/request memory, without application persistence or request-body logging. Documents reach the server when you generate; selected passages reach the provider. Hosting infrastructure and providers have their own data policies. OpenAI storage is disabled in the request; that does not promise zero retention. Clearing the session drops application references; it cannot erase provider records or guarantee memory zeroization. Shared rate and concurrency limits apply per server instance; platform-wide abuse protection is a deployment concern.
Explore the PagePilot case study →