Retrieval you can
read back.
AskDocs answers from your archive, shows the sources it used before it writes a word, and runs the whole pipeline on hardware you control.
Upload. Index. Ask. Verify.
Four steps, and the fourth is the one that matters. Every answer lands with the documents it came from attached.
Upload
Drag in files or whole folders. Duplicates are detected on content, so re-uploading a folder does not double your archive.
Index
Documents move through pending → indexing → ready. Small text files finish in seconds; a long PDF can take a minute.
Ask
Plain language, across everything you have access to. No query syntax to learn and no folder to remember.
Verify
Open the source. Every claim carries a numbered citation with the document, the page, and the passage it rests on.
Sources arrive before the answer.
This is an ordering decision, not a layout one. The retrieval step streams its sources first, so you can see what the answer will be built on before a single sentence is generated.
Each source shows the citation number, the document title, where in it the passage sits — page, slide or sheet — a preview of the text, and how strongly it matched.
If the context does not contain the answer, AskDocs says so directly. It does not speculate.
That rule is in the system prompt, not in the marketing. AskDocs' job is to point you at the right source, not to replace reading it.
Two searches, fused.
Meaning-based search finds the passage that answers your question in different words. Keyword search finds the exact term you typed. Running only one of them is how retrieval misses.
-
Dense search
Vector similarity over locally generated embeddings, indexed for fast nearest-neighbour lookup. Catches paraphrase.
-
Full-text search
PostgreSQL full-text search over indexed columns, in the same database. Catches the literal term, the product code, the surname.
-
Fusion
The two result lists are merged by rank rather than by score, so neither search has to be calibrated against the other.
-
Rerank
A cross-encoder re-reads the shortlist against your actual question and reorders it. Skipped for very short queries, where it adds latency without adding signal.
-
Answer
Only the top passages reach the language model — with their citations already attached.
The files you actually have.
Including the scanned ones. Images and image-only PDFs go through OCR automatically — no separate step, no separate tool.
| Format | Handled as |
|---|---|
| Text extraction, with OCR fallback for scanned pages | |
| DOCX | Document text, headings preserved for citation context |
| PPTX | Slide text, cited by slide number |
| XLSX | Sheet contents, cited by sheet |
| MD · TXT | Read directly |
| Images | OCR |
Who can see what, on the record.
Workspaces
The isolation boundary. Documents, conversations and indexes belong to one workspace and are never pooled across them.
Roles and ACLs
Viewer, member, admin and owner, plus per-document permissions granted to a role, a person or a group. Retrieval respects them before it searches, not after.
Audit log
Append-only. Every upload, deletion, permission change and membership change is logged, with who did it and when.
Your models, your call.
Embeddings run locally on CPU by default. No API key, no GPU, no third party involved in turning your documents into an index.
The chat model is chosen independently and swapped with a single environment variable. Hosted endpoints and local runtimes are both supported, so the same deployment can move from a hosted model to a machine on your own network without reindexing anything.
If you point AskDocs at a hosted model, your question and the retrieved passages go to that provider. That is worth knowing before you choose one.
Run a local model instead and nothing leaves your network at all. See self-hosting for what that takes.
What it will not do.
These are positions, not gaps. Knowing where a tool stops is what makes the rest of it trustworthy.
Not an editor
AskDocs reads documents. Editing stays where it already happens.
Not enterprise search
Tuned for small and mid-sized teams. Past about a million documents, you want a different class of tool.
Not a web assistant
If the answer is not in your documents, it says so rather than reaching for the internet.
Not auto-action
It retrieves and cites. Deciding what to do about it stays with you.
Run it yourself.
Four commands, Docker, and a machine you already own.