Astrallabe
Get started

About

An answer is only worth its source

Astrallabe started from a simple observation: checking an answer often takes longer than searching on your own. The product is that correction.

A conventional assistant

It answers from memory

It produces a plausible answer from what it learned, and invents a reference when one is missing. Nothing on screen separates what it knows from what it assumes.

“According to a 2021 study, the effects persist beyond three years.”source not found

Astrallabe

It answers from your documents

The useful passages are found in your corpus, given to the model, and the model must cite the ones it uses. No passage, no answer.

“Follow-up stops at eighteen months.”Moreau · DOCX · p. 41

How it works

Four steps, always the same

There is no alternative path: a question follows this route, or it gets no answer.
  1. 01

    Store

    PDF or Word. The text is extracted; the original file is kept. A link pasted into the question is read for that answer only.

  2. 02

    Split

    Into passages of about two thousand characters, without cutting a sentence.

  3. 03

    Retrieve

    Two engines in parallel, merged into a single ranking.

  4. 04

    Cite

    Document, page, literal excerpt. Numbered in the answer.

The scope we claim

Three commitments

Three promises kept at every question: visible sources, no invention, and your files that stay yours.
  • A summary is always cited

    Ask for a summary of any document: it is built part by part, and every sentence points to the pages it comes from.

  • No answer without enough source

    When your documents are not enough, Astrallabe says so and shows what would need adding, rather than filling the gap.

  • Your corpus trains no model

    Your documents serve to answer you, you alone, and nothing else.

Technical sheet

The exact numbers

Every value is a documented trade-off. A shorter passage would lose its context; a longer one would dilute the topic and the search would lose precision.
Passage size
≈ 2,000 characters
Overlap between passages
300 characters
Vector dimensions
1,024
Vector index
HNSW, cosine distance
Full-text search
PostgreSQL, French dictionary
Merging the two rankings
Reciprocal Rank Fusion, k = 60
Passages injected per answer
6 at most
Data isolation
Row Level Security

Processing in Europe

Storage and compute in Paris, for indexing as well as for generating answers. A web search or reading a video sends a request to DuckDuckGo or YouTube, not your documents. No payment data passes through our servers.

Deletion

Deleting a source also deletes its original file, its passages, its vectors and the citations that referred to it, in the same operation. The export is a ZIP archive containing the account JSON, the sources with their passages, the conversations, and the original PDF or Word files. The original file is kept in the European Union, accessible to its owner alone, deleted with the source or the account, and included in the export.

Try it on your own documents

It is the only way to judge: drop in a report you already know, and check the cited passages.

Create my account