Skip to main content
When you ask the agent a question about your files, it does not read every document. It searches, retrieves the few passages that matter, and works from those. The quality of that search decides the quality of the answer, so it is built with some care.

​
From upload to searchable

1

Queued

Every document that enters Casa Conect is queued for indexing, whether it was uploaded, filed from email, synced from a drive or added to a Shared Space. A document that is being edited continuously is indexed once the edits settle: about two minutes after the last change, and at least every ten minutes while editing continues.
2

Scanned

The file’s malware scan must come back clean before it is read.
3

Parsed

Text, tables and layout are extracted, including from scanned pages. If the first pass produces a low-confidence reading, typical of a poor scan, the document is automatically re-read with a more capable parser.
4

Divided into passages

The document is split along its own structure. Headings, tables, lists and figures are natural boundaries, and each passage remembers the heading it sits under. Long text is divided at sentence boundaries by a splitter that understands Spanish legal writing. Abbreviations such as S.L., N.I.E., D.N.I. and C.I.F., and ordinals such as 1.º and 2.ª, do not break a sentence in two.
5

Indexed twice

Each passage is indexed for meaning, as a vector embedding, and for words, as a full-text index. Its language is detected so that the right linguistic rules apply.
Most documents are searchable within a few minutes. Large scans take longer. Until then the agent may tell you that a document is not indexed yet, and Populate waits for indexing at its Waiting for sources stage.

​
How a search works

Before anything is searched, Casa Conect works out exactly which documents you may read from your vaults, your projects and your permissions in each Shared Space. The search runs only inside that set. A narrower filter can reduce it, never widen it.
Your question is reduced to the terms that matter: names, identifiers such as a NIE or a cadastral reference, amounts and key words. Filler words are dropped. If this step fails, your question is used as it was written.
A semantic search finds passages that mean something similar, even in different words. A keyword search finds exact terms, which matters enormously for names, numbers and references that a meaning-based search handles poorly.
The two result lists are merged in a way that rewards passages both searches agree on, and a shortlist of candidates goes forward.
A dedicated reranking model reads each candidate against your question and reorders them. The best few are returned.

​
Spanish and English

Search is bilingual by design.
  • Each passage is tagged as Spanish or English, defaulting to Spanish when in doubt.
  • Keyword search uses the right stemming for each language, so arrendamiento finds arrendamientos. It also ignores accents, so hipoteca finds hipotéca as typed by a hurried colleague.
  • Every query is run against both languages, and the better match wins. A Spanish question therefore finds an English contract, and the reverse.
  • Spanish identifiers such as Y-7654321-K and 44.567.891-K are kept intact instead of being split at their punctuation.

​
What the agent sees, and what you see

A search returns excerpts, each with its section title, the document it came from, its scope and a relevance score. The agent then opens the documents it needs in full. In the conversation, a search appears as “Searched your documents” with the query it used. Elsewhere in the product, values extracted from documents carry citations:
  • In a Shared Space Summary, each AI-confirmed value has a numbered citation. It opens a preview of the source with the relevant passage highlighted, and an Open in file viewer button.
  • In the editor, values placed by Populate carry a source marker showing which vault the source is in, and the excerpt.
If a citation points to a document you are not allowed to read, you are told that a source exists, but you cannot see its contents.

​
Getting better results

​
What is searched

Deleting a document removes it from the index.

​
Services involved

Parsing, embedding, query understanding and reranking are performed by specialist services, listed with exactly what each receives under AI data handling. The index itself, both passages and vectors, is stored in Casa Conect’s own database in the EU.