Knowledge sources
Deep document analysis
Turn tables inside documents and images into data, build an entity index across files, and choose the models that do the work.
Документация доступна на английском и турецком, как и интерфейс Harmona. Эта страница показана на английском.
Agents can always read your documents in full. Two optional steps go further: deep document analysis turns structured content into tables, and the entity index records what each document is about. Both run in the background and never slow down chat.
Open a connection and click Document Processing. You see three tabs: Deep Document Analysis, Entity Cards and Spreadsheet tables.
Deep document analysis
Menus, price lists and catalogs often sit inside PDFs, Word files or photos. Deep analysis transcribes that content once into a clean data table. Agents can then do accurate math across files, for example compare prices across all menus, instead of picking numbers out of text.
- Auto on: every document and described image added from now on is analyzed once. Files already in the system are not analyzed automatically.
- Auto off: nothing is spent. Documents stay readable as text. You can still click Analyze on any single file.
Each file shows its result: Table extracted, No table found (the file is used as cited text), Not analyzed, Analyzing... or Failed. A failure is often a scanned PDF with no readable text. Partial means the output was cut short and the table may be missing rows at the end. Click Re-analyze to transcribe a file again and replace its table.
Cost: one AI transcription per file, once. Results are cached; a changed file is analyzed again. The cost is billed to the connection, not to an agent. Pages synced from ReadMe are not analyzed automatically; use Analyze on a page when you need it as data.
Entity index
The entity index records what each document breaks down by: its cities, products, counterparties or periods. Questions like "list all X" or "X grouped by Y" then get complete answers, and searches go straight to the right files.
Each document gets a small card. On the Entity Cards tab, click Generate N missing cards to index files that do not have one yet. Spreadsheets are indexed separately, and images or scanned PDFs need document analysis first. Cost: one small AI call per document, cached until the file changes.
Turn it on
Both are organization settings under Management → Organization Settings → General, in Connection Settings. Both are off by default.
- Auto deep document analysis: transcribe new documents into tables.
- Automatic document entity indexing: give every new document an entity card.
The Document Processing button on a connection shows whether each is on (Analysis on, Index on) or off.
Внимание
Analysis models
When model choices are available for your organization, the same settings page lets you pick the model for each background job. Options marked with a key use your organization's own API key.
| Setting | What the model does |
|---|---|
| Analysis model | Transcribes documents into tables, at upload or when someone clicks Analyze |
| Indexing model | Writes entity cards. A small, fast model is recommended |
| CSV & Excel analysis model | Describes uploaded spreadsheets: column meanings and structure |
| SQL analysis model | Describes connected database tables |
| SAP HANA analysis model | Describes SAP HANA tables, which are often larger |
The main agent model still writes the answers. These models only prepare your data.
Обновлено 2026-09-24