Supported Upload Formats

Your collection is rarely PDF only: Word reports, execution spreadsheets, presentations, pages saved as HTML — and all of it often zipped together. The Upload page accepts Office documents (DOCX, XLSX, PPTX) and HTML as files, including inside ZIP/TAR/GZ packages. Each format is extracted preserving its structure (sections, spreadsheets, slides) and follows the same path as always: splitting into excerpts, semantic vectors, entity identification, graph and search. You upload what you have, the way you have it.

This guide details what is accepted, the limits and the rejection rules. The general upload flow — target context, batch queue, tracking — is in the upload and ingestion guide.

1. Accepted formats

A file outside the list is refused with:

"Formato de arquivo não suportado: .<ext>. Formatos aceitos: PDF, DOCX, XLSX, PPTX, HTML, TXT, CSV, JSON, XML, MD, TSV e LOG." ("Unsupported file format: .<ext>. Accepted formats: PDF, DOCX, XLSX, PPTX, HTML, TXT, CSV, JSON, XML, MD, TSV and LOG.")

2. Limits per format

These are the default values. The administrator can adjust them at installation through the environment variables listed below.

Exceeded the format's limit? The message is:

"O arquivo "<nome>" excede o limite de <N> MB para o formato <FORMATO>. Divida o conteúdo em arquivos menores e envie novamente." ("The file "<name>" exceeds the <N> MB limit for the <FORMAT> format. Split the content into smaller files and upload again.")

Why do DOCX and PPTX have a lower ceiling? Because reading those two formats loads the file into memory. XLSX uses streaming (SAX) reading and therefore accepts much larger files without risking the service's memory.

3. Rejection rules (security)

All validation happens before the file touches the extractor, and the refusal comes back as a validation error (HTTP 422) with a message in Portuguese — never a technical stack trace. The messages below are the real system messages:

How it works under the hood: validation checks the extension, then the limits, then the magic bytes (%PDF, the PK.. ZIP signature of current Office formats, the OLE2 signature of legacy Office) and finally the Office container's content — [Content_Types].xml plus the main part (word/, xl/ or ppt/), inspecting at most 500 entries. The global anti zip-bomb protections for Office formats are a minimum inflate ratio of 0.01, a maximum entry of 256 MB and at most 10,000 entries — all configurable through the DATTA_OOXML_* variables.

4. How each format becomes sections

The extracted text is split into sections, and it is the sections that guide the split into excerpts and the citations search returns.

After extraction, nothing changes compared to the older formats: splitting into excerpts (200 words with a 30-word overlap), automatic semantic vectors, entity identification with the context's model, writing to the graph and to the indices, full lineage — and the original file is stored with a SHA-256 hash, recoverable on the document's screen.

Entity identification in long documents

Entity identification no longer truncates the text at the first 10,000 characters. Documents larger than one window (DATTA_NER_WINDOW_CHARS, default 40,000 characters) are split into windows aligned on paragraph breaks and analyzed with controlled parallelism (DATTA_NER_CONCURRENCY, default 2 — the low number protects the AI provider's per-minute quota). The entities and relationships from the windows are merged with deduplication: the same type plus the same name yield a single entity, and complementary properties are merged.

AI cost is capped per document: at most DATTA_NER_MAX_WINDOWS windows (default 24, the equivalent of ~960 thousand characters). If the document has more windows than that, they are distributed evenly across the beginning, middle and end, and the file ends with the warning "NER cobriu N de M trechos do documento (amostragem distribuída…)" ("NER covered N of M excerpts of the document (distributed sampling…)"). Raise DATTA_NER_MAX_WINDOWS when full coverage is worth the cost.

5. Packages (ZIP/TAR/GZ)

The new formats also work inside packages, in batch upload: each processable extracted file joins the same queue in the executions hub. Malicious paths (zip-slip) are ignored and the total decompressed size is limited to 500 MB per package.

Practical example — a ZIP with a report, a spreadsheet and slides

1. You have prestacao-contas-2026.zip with relatorio.docx (12 MB), execucao.xlsx (40 MB) and apresentacao.pptx (18 MB).
2. In Processar›Upload, select the target context and upload the ZIP.
3. The platform unpacks it safely and each file joins the same processing queue, visible in the batch's group.
4. At the end: the report became sections by heading, each spreadsheet tab became a section with the rows as displayed, and each slide — including the presenter notes — became a section of its own.
5. Search for a term from the report: the result cites the source section, and the original file remains available on the document's screen.

6. Troubleshooting