CITEPOST

Choose a synthetic manual, contract or policy corpus, or upload a text-based PDF. The lab records the real ingestion stages, retrieves lexical and vector candidates, combines their ranks and runs a local reranker. The inspector exposes the selected passages, index snapshots, timing and GroqCloud token usage.

Every generated factual segment must cite a source from the actual context. Click a citation to open the versioned PDF at the source page, with the cited text highlighted beside it. An evidence gate can refuse weak retrieval; unknown questions and controlled failure scenarios are part of the demonstration.

The public lab has temporary isolated sessions, three PDF uploads per session, a 5 MB/60-page file limit and two-hour retention. The interface and local retrieval models are English. Parsing and structural chunking are heuristic; OCR, pixel-level PDF highlighting and token streaming are not included. Citation validation checks provenance, not factual truth, and prompt injection remains a risk.

The public repository includes a separate tuning set, 60 held-out evaluation cases and measured results. This is an owned synthetic demonstration, not a client deployment or a promise to eliminate hallucinations.