# Documents

The Documents section of a Knowledge (RAG) or SharePoint pool lists its documents and is where you upload new ones.

PHOENIQS Data Pools: Documents
PHOENIQS Data Pools: Documents

# Uploading

Drag files onto the upload area, or click it to choose them. Several files at once work.

Accepted
Documents PDF, DOCX, PPTX, XLSX
Text Plain text, Markdown, CSV, JSON, YAML, XML, HTML
Scans and images PNG, JPEG, TIFF, and PDFs without a text layer, read by OCR

Each file can be up to 1 GB, and uploads count towards your Data Pool's storage. Files are checked by their content, not their name, so a file whose content does not match an accepted type is refused.

In a SharePoint pool there is no upload area: the files come from the SharePoint site. See Pool types.

# The document list

Column What it shows
File The file name. In a SharePoint pool, it links to the original.
Status Queued, Processing, Ready or Failed, updated by itself while files are processing. Ready · OCR means Data Pools read the text from an image or scan itself, by OCR. low confidence means the OCR result may be incomplete; see below.
Chunks How many passages the document was split into for search
Size The file's size
Actions Details and Delete

# How documents are read

When a document is uploaded, Data Pools:

  1. Converts it to text. PDF, DOCX and PPTX files are converted by Docling, an open-source document converter. It keeps headings, lists, tables and reading order, and reads scanned pages with its built-in OCR. Spreadsheets keep their sheets and tables. Text, CSV and HTML files are read as they are.
  2. Describes pictures. Diagrams, photos and screenshots in PDF, DOCX and PPTX files get a short description from an AI vision model, so what they show can be found by search. Small pictures, such as logos, are skipped.
  3. Reads images. PNG and JPEG files are transcribed by an AI vision model, with a built-in OCR engine as a check that nothing was missed. TIFF files are read by the built-in OCR engine.
  4. Splits and indexes the text into passages, searchable by meaning and by keyword.
  5. Describes the document. An AI model writes a description and extracts metadata: the people, organisations, topics, locations and timeframe it covers.
  6. Updates the pool's summary, the overview of the pool's contents that agents read first.

The first four steps decide when a document is Ready and searchable. The description and metadata follow in the background and never hold up an upload.

If Docling cannot convert a PDF, Data Pools reads it itself. Pages with a text layer are kept as they are, and scanned pages are transcribed by an AI vision model. Invoices, orders and forms with a text layer come out as structure: fields (such as invoice number, date or IBAN), line items and totals, alongside the plain text. Agents can ask for this structure directly.

# AI models used

Step Files Model When it runs
Describing pictures PDF, DOCX, PPTX Qwen3-VL-235B (inference-qwen3-vl-235b) Each picture that covers a meaningful part of a page
Reading images PNG, JPEG DeepSeek-OCR (inference-deepseek-ocr) Every image
Reading scanned pages PDF DeepSeek-OCR (inference-deepseek-ocr) Only when Docling cannot convert the PDF
Reading invoices and forms PDF with a text layer Llama 4 Maverick (inference-llama4-maverick) Only when Docling cannot convert the PDF, for invoices, orders and forms
Indexing for search All BGE-M3 (inference-bge-m3) Every passage, and every search question
Descriptions, metadata and pool summary All GLM 5 (inference-glm5) After the document is Ready, in the background
Ranking search results All BGE Reranker (inference-bge-reranker) Every search

All of them run on the PHOENIQS AI Model Service in Switzerland, which does not use your data to train models. Details on a document shows how it was read. When a model changes, the changelog says so.

# Confidence

Every image or scan that Data Pools reads itself gets a confidence between 0 and 1. Below 0.6, the document is marked low confidence in the list, and agents are told so too. Check those documents: a poor scan, handwriting or an unusual layout can leave gaps in the text.

# Document details

Details on a document shows:

Row What it shows
Status, Chunks, Size As in the list
Description The AI-written description agents receive
Metadata The people, organisations, topics, locations and timeframe found in it
Extraction How the text was obtained, and with which model
Structure The fields, line items and totals found in a scan or invoice
Confidence How complete the OCR reading is likely to be
Pages The number of pages and how long reading them took
Sheets For a spreadsheet: how many sheets were indexed, and its tables, records and charts. Hidden sheets are not indexed.
Document type For a scan or invoice: the kind of document and its language
Extraction notes Anything that went wrong while reading the document

# Deleting a document

Delete removes the document, its passages, description and metadata at once, and agents stop finding it immediately. It cannot be undone.

# Test retrieval

Test retrieval shows what a search of this pool returns, before you rely on it in an agent. It is visible to administrators only. Type a question in plain language and click Search. Each result is a passage, with its chunk number and a relevance score.

If the right passages do not come back:

  • Check the document is Ready and not low confidence.
  • Phrase the question the way the document does. Searches combine meaning and keywords, so the document's own terms help.

Tables
../tables/
Chat and agents
../chat-and-agents/