Document Intelligence

When a user uploads a file, Aiguro automatically processes it for optimal discoverability - regardless of the original format (PDFs, Word documents, spreadsheets, CSVs, and more). The content is prepared and indexed so the AI can search and reason over it semantically, and relevant signals are surfaced as discoveries automatically. For spreadsheet-style documents, Aiguro can also analyse the underlying structured rows and columns directly, which helps it produce more quantitative, evidence-based discoveries.

Spreadsheets and tabular files additionally receive automated quality analysis: Aiguro inspects the structure of each sheet and flags common issues that could affect data accuracy - such as headers not starting at the first row, leading summary rows, merged cells, or multiple separate tables on the same worksheet. Where issues can be safely corrected automatically, Aiguro applies the fix and notes what changed (including any data that may have been excluded). Users can view the full quality report directly on the Data Sources page by clicking the info icon next to any analysed document. For unusually large spreadsheets, Aiguro uses a lighter-touch check to keep processing fast, and the quality report notes when this happened - the data itself is never trimmed. Spreadsheets beyond the per-file import limit (measured in cells, i.e. rows x columns across all sheets) are not imported at all: the quality report names the affected sheet and the limit so the export can be split into smaller files and uploaded separately. Many findings can be corrected right in Aiguro, without editing and re-uploading the file: the import settings of a spreadsheet show its first and last rows as they will be imported and let users with write access choose the header row (including two-row headers such as "Revenue" over "2024 | 2025"), rename or skip columns, set a column's type (number, currency, percent, date or text, with its decimal separator or date order), leave out rows or whole sheets, and keep total rows the automatic check would remove. Applying the settings re-imports the unchanged original file, and the settings are kept for every later re-check.

Aiguro also recognises when the same file is uploaded more than once. If an identical document is added to several workspaces, the heavy work of preparing and indexing it is done only once and the result is safely shared, saving time and processing cost. The shared copy is automatically cleaned up once it is no longer used anywhere.

While a document is being prepared, the Data Sources page shows which step it is in - waiting, converting or indexing - and how far indexing has progressed, updating on its own. For very long documents that already contain readable text, Aiguro skips reading text inside embedded images to keep processing fast; the document then starts with a short note saying so, and scanned documents are always read in full.

Users can add their own short description to any uploaded document - for example, what it's for, what time period it covers, or which project it relates to. This context isn't just a note for humans: it's fed to the AI when generating discoveries and when agents search or list documents, so the added context sharpens both the signals Aiguro surfaces and how well agents can find and reason about the right document.

Aiguro also keeps a living set of analysis notes per document: each time a document is analysed, a short dated note records what that analysis covered. When the document is re-analysed later, these notes steer the AI toward new angles instead of restating insights it already found, so repeated analyses keep producing fresh discoveries.

Re-analysing a document is only worth it while it still has something new to say. When several analyses in a row find nothing new - or the AI sees that a document clearly has nothing to do with the workspace's business area and focus - Aiguro marks it with a quiet "No new findings" hint. For a long document, which a check reads a part at a time, only complete passes through the whole document count, so it is never marked for parts nobody has read; a check focused on a topic reads only the matching parts of such a document and does not count either. Automatic and scheduled checks then skip it, while chats, table questions and the knowledge graph keep using it as before. A manual check still looks at it and removes the hint as soon as it finds something new; editing the description or choosing "Include in checks again" also brings it back. Separately, documents that nobody has used in chats or findings for more than 90 days show a "Rarely used" hint, so teams can see which files may no longer be needed.

On the Data Sources page, users can also organise documents into groups (with an Ungrouped bucket) to keep related files together while preparing and reviewing discoveries.

A document can be made available in other workspaces without uploading it again: in the share dialog, users pick one or several workspaces they can edit in one go, workspaces that already have the file are marked, and afterwards they see for each workspace whether sharing worked. Only the file is shared - analysis results, findings and data fields stay separate per workspace. Teams that set up several workspaces with the same structure can also copy a whole group into other workspaces: each chosen workspace gets a group with the same name containing the same files. The copy is independent from then on, so later changes to either group stay in their own workspace.

Processed documents also expose a Details page where teams can review extracted content, quality findings, document grouping, and mapped data fields in one place. Everyone with document access can open that detail view, while editing controls remain available only to users with write permissions.

Spreadsheets can also feed data fields - curated dimensions such as supplier, material group or cost centre that users later filter, chart and select values from. Setting these up used to be entirely manual. Now, once a spreadsheet has been processed, Aiguro proposes which of its columns could become a field, either by matching an existing field (for example when a second file contains the same supplier list) or by suggesting a new one, each with example values and a short explanation. Where earlier analyses have verified that a column in one file corresponds to a column in another, that relation feeds the same proposals, so a mapping made in one file is suggested for its counterpart in the other. Suggestions appear in the data center's field area, and a small hint next to each processed document points to them. Nothing is applied on its own: the user accepts or rejects every suggestion, and a rejected column is not proposed again.