2 · configure
3 · deploy
AVAILABLEAvailable product — files, demo URL and secure download get mapped in EDF 0.3.
Answers only from your documents.
A help assistant that answers only from your documents, shows where every answer comes from, respects who may see what, says 'I couldn't find this' instead of guessing, and tells you what people ask that you have not written yet. Ingest Markdown, HTML, text or FAQ JSON; retrieval combines a BM25 keyword index (typo correction, synonyms, phrase bonus) with an offline vector by reciprocal rank fusion, only over documents the caller may see through per-document access groups. A confidence score decides whether to answer; answers are extractive (the best cited sentences, no language model, works offline) or written by your own model from a fenced context and then checked sentence by sentence against the passages they cite. Documents that contain wording aimed at the assistant are held back until a person approves them. The dashboard has an Ask page that shows confidence, sources and freshness, Sources with usage and staleness, a Gaps report that groups what people asked and could not get answered, Quality with a golden-set evaluation and a confidence-threshold sweep, Analytics, settings with a synonym editor, API keys with groups, and a drop-in website widget. Node.js 22.5+ and SQLite, no npm dependencies, 53 automated tests. The offline vector is lexical (typos and word forms), not semantic: the guide says plainly what it misses, and the shipped evaluation (36 questions on a demo help centre, tuned by the developer, one failure kept on purpose) is a regression test, not a benchmark.
Origin: EDF native product (Platform Modules 1.0) · Roadmap window:


Placed in the private vault layout (deploy/vault/products). Downloads stay disabled until checkout is connected.
Derived by the mapping engine (family · layer · tags · dependency rules). Stored in the catalog as compatibility, dependencies, relations[].
| VERSION | STATE | NOTES |
|---|---|---|
| v1.0.0 2026-10-13 | AVAILABLE | Initial release. This package contains: Ingest Markdown, HTML, plain text and FAQ JSON; passages split along headings with heading paths and overlap; unchanged documents skipped, changed ones replaced; Hybrid retrieval: BM25 with typo correction, synonym groups and a phrase bonus, fused with an offline hashed vector by reciprocal rank fusion; optional OpenAI-compatible embeddings; Access control per document: access groups on documents, groups on API keys; restricted documents never appear in search, answers, suggestions or citations (tested); Answers: extractive with numbered citations, or your language model from a fenced context with every sentence checked against the passage it cites; falls back to extractive and says so; Refuses below a confidence threshold, with suggestions and a logged gap; follow-up questions rewritten from history; spelling corrections shown; Safety: documents with wording aimed at the assistant held back for review (including text hidden in HTML comments); questions logged with e-mail, numbers and credential URLs masked; answers never stored; retention and purge. |
| → | Roadmap | Target window for the next stage. |
Standard (personal or commercial use in one end project or your own business) · Extended ×3 (SaaS, unlimited client projects, finished products only). Agency licence comes later.
Activated by 1.0 (download) and 2.0 (one-click deploy). Download will be a 10-minute signed URL from the R2 vault — never a public /files/product.zip.