Documents Automation

Use Documents Automation, after Connectors in the sidebar, to work with Connectors, Documents, Templates, Review and Monitoring in the left navigation. Choose Hide navigation to give the document viewer more room; Show navigation brings the sections and document tree back. Folder discovery, original previews, managed uploads, versioned templates, extraction, human review and recurring scans are available. Actual accuracy depends on your documents and templates; test representative examples before production use.

Connect a folder

1. Sign in and select the intended tenant and project.
2. Open Documents Automation → Connectors, then choose Connect folder. Choose any available folder connection, including a connector-backed folder or S3. Unsupported document connections are marked.
3. Enter a display name. Leave Relative folder empty to use the connection's root, or enter a subfolder such as invoices.
4. Choose Connect folder. AgentData starts an inventory scan. Expand Show folders and scan settings to see each connection, last scan time, observed file count and scan method. A partial scan is marked; its count is not a complete total.

Administrators can choose New local folder connection in the same dialog. Setup proceeds through Name, Connect, Install when needed, and Review, with Back/Next navigation and persistent Close controls. For a Mac folder, select Connector (local folder); Local means the AgentData server's filesystem. See Connect a folder on your computer for installation and token guidance.

Editors can attach authorized existing folders and scan them. Administrators create new connections and manage processing consent. Viewers can browse authorized documents. If a connection is missing, check your tenant, project and permissions with its administrator.

Upload documents

In Connectors, expand Upload documents, create an upload folder and upload supported PDF, PNG or JPEG originals. Processing consent must be enabled first. Managed uploads are separate from connected customer folders: uploading does not write into the folder on your computer. Uploads and retained artifacts expire under your workspace's retention policy.

Browse, parse and create a template

The filename appears in the compact page header. Process document opens one workspace for the brief, labels, discussion and results. View gives the original more space. Page navigation, zoom, settings and download remain in the viewer toolbar. Drag either panel divider to adjust its width. Move the pointer to the image edges for previous/next document and folder navigation; use the wheel to zoom and View mode to pan.

1. Expand a folder in the left panel and click a filename. Its Info icon opens properties. PDF pages are rendered as images; Download original retains the source format.
2. Choose Process document. Opening the workspace does not start OCR or an AI call. In Brief, describe the document and what to extract, optionally choose a saved template, and review the document settings. Enter a brief or choose a template before proceeding.
3. Choose Next: optional labels. You may skip this step. To guide field mapping, draw a rectangle around an example value and give it a label, output key and data type. Select an existing label to rename or remove it. For tables, draw an example repeating row with its fields inside it.
4. Choose Save and continue / skip labels, select a reader in Test and compare, then choose Run extraction / apply feedback. The selected reading engine prepares the document; AI maps the content into the requested fields and rows. Rescan and test forces a fresh reading pass. Progress, errors and processing details remain visible.
5. Inspect Results → JSON or CSV / Table, compare values with the source and refine the instructions in Test and compare. Enter a template name in Results and choose Save template to save a reusable, versioned definition. Trial results still require review before publication.

Briefs, labels and discussion changes autosave to your private document draft. Save labels and draft also saves explicitly. Wait for the saved confirmation before leaving or reloading. A failed save remains visible; resolve it and save again. Closing the side panel preserves the current draft, and saved labels reload when reopening the document. Saving a reusable template is a separate action.

Green indicates completed reading, red indicates not processed or failed, and amber indicates queued, running or partial processing. A green status does not certify accuracy. Labels identify sample regions and desired fields; they are not verified expected values or OCR training. You can label before the first run, or add/refine labels after inspecting a result and run extraction again. Regions remain tied to the original sample/page; a different document is not assumed to have the same coordinates.

Templates support text/identifiers, integers, exact decimals, dates, booleans and configured enumerations. Runtime inputs supply context or validation settings; they are not source evidence.

Label a repeating row

For lists such as blood tests or invoice items, choose Process document → Labels (optional) → Draw repeating row, then draw one outer rectangle around a complete example row. Name it, for example Blood tests. Choose Draw field and draw smaller rectangles inside for the test name, result, units, range and date. Each enclosed field joins the row automatically. Its Repeating row dropdown lets you change that assignment or keep it as a document-level field. Fields already drawn inside a new row rectangle are also assigned to that row.

The row creates a repeating output table and each field creates a column. Use clear field definitions and the Brief to describe shared values, such as applying a report date to every test. Choose text for results containing signs such as <0.5; numeric companion columns can be added later. The example guides extraction of all matching rows and pages, including rows outside the rectangle. It is not a crop, a copied value or a guarantee of recognition accuracy. Remove a row with Remove repeating row; its fields remain as document-level fields until reassigned or removed.

Reuse label names

In the label editor, type a dotted name such as claim.amount or claim.amount.net, or choose a previously saved name from the field's dropdown. Choose Save reusable label name to make the name available across documents in the current project. The dots define a naming hierarchy; JSON keys and CSV columns remain explicit mappings and are not automatically nested JSON objects.

Expand Reusable label names to add a path or remove one. Select Also remove names below this path to remove a whole subtree. Removing a catalog name does not delete rectangles or rewrite saved templates. Use the image label's × to remove a label from the current draft.

Document language and scan settings

Click the small Info icon beside a filename to open Properties. Clicking the filename itself opens the original document. The dialog shows its source path, modification time, last scan, size, format and processing status, followed by language and recognition settings. Hover the filename for a short summary. For several documents, tick their checkboxes (or Select all loaded files), choose Properties, change the desired settings and choose Apply changes. This saves without starting recognition. Only settings you change apply to all selected files; their other properties remain separate. Each file reports whether its update succeeded. Selection covers loaded files in the active folder, not unloaded inventory pages.

Language, quality and content properties are available in Process document → Brief beside the extraction goal. Recognition mode is selected with the reader in Test and compare. You can also open Document settings (⚙) for the viewed file and choose Save document settings. Neither action starts processing. Enter the extraction goal and optional labels first, then run a test in step 3. A spinner shows processing is running; completion, partial coverage or a failure is shown explicitly. The language picker is a searchable dropdown with checkboxes: search by name or code and select several languages. The available languages depend on the installed OCR packs; English and Hebrew are configured initially. Language selection guides recognition; it does not translate the document.

• Automatic — embedded text + OCR when needed is recommended for most files. The parser uses embedded PDF text and also runs OCR on scanned or image-containing pages. This is a parser rule, not an LLM decision. Always OCR reads each page image even when embedded text exists; use it when embedded text is garbled or incomplete. The later extraction step maps the reading result into the requested fields. This recognition-mode setting applies to Local and Docling readers; Vision reads page images directly.
• For the Local reader, Document quality: Automatic uses original-image contrast and dimensions to identify low contrast or small images. When enhancement is enabled, it adjusts preparation and upscaling within processing limits. Poor forces enhancement and bounded upscaling; Good keeps your selected OCR settings. Upscaling does not recover missing detail. In Brief, Detected image quality shows per-page dimensions and reported DPI when available after scanning. Older scans need rescanning to record these diagnostics. The rating is a heuristic, not a test of OCR accuracy, blur, handwriting or language; review and override it if needed. Embedded-text pages do not need an image-quality rating.
• Detect tables enables the Local reader’s native PDF table detection or Docling’s table-layout recognition. Vision preserves rows for subsequent field mapping.
• Enhance low-contrast scans applies contrast and denoising before Local OCR. It is enabled by default for new settings. Previously saved choices remain unchanged, including an explicit opt-out. The original file is preserved. Turn enhancement off and rescan if it makes recognition worse.
• Contains handwriting forces OCR and marks the parse for human review. Contains meaningful images also marks coverage incomplete for review. Accuracy depends on the selected reader and source quality; these flags do not guarantee handwriting or visual interpretation.

Settings follow the same connected folder and file path across inventory scans. They affect new parsing runs; existing results retain their original settings. Save the settings, then open Process document → Test and compare and choose Rescan and test. This rereads the document and reruns structured extraction; it does not approve values. A stale concurrent edit shows a conflict instead of overwriting another user's settings.

Choose a reading engine

In Test and compare → Reading engine, choose an available reader:

• Local parser / OCR uses embedded PDF text and the installed OCR engine.
• Docling service reads original pages and table layout, retaining cell positions where available. Availability depends on the configured service; a sleeping service can take longer to respond.
• Vision model sends original page images to the configured image-capable AI model. It first reads rows and text, then the shared extraction step builds typed JSON/table fields. Its evidence points to a page rather than an exact cell, so verify values against the original.

Unavailable engines are disabled. There is no silent fallback to another service or model. Document-processing consent and workspace data-access rules still apply. Docling hosted in the cloud is cloud processing; later AI mapping may use a separate provider. Labels help identify the desired structure and locations, but cannot recover unreadable pixels. Docling reads the layout first; prompts and labels guide the later mapping. Vision also receives the brief and applicable sample labels during image reading.

Move between preparation and testing

Test and compare replaces the main labeled-document viewer with reader controls and the JSON/CSV result area. Choose a reading engine and Run extraction / apply feedback to start; simply opening the step does not process the document. Brief and Labels (optional) return to preparation with your draft and source labels retained. View opens the original again.

If a draft save fails, a message appears next to the step buttons. You can still return to Brief or Labels and correct it; running a test waits for a successful save. JSON field keys must be unique lowercase English identifiers such as test_result; display labels can use Hebrew or another language. Correct the input and save again, or choose Retry save for a temporary failure.

If another window changed the saved draft, AgentData does not overwrite it. Load saved draft asks before replacing your unsaved edits with the latest saved version. Choose it only when you want to keep that saved version. A saved message applies to a completed save; new edits must finish saving before you close or reload the page.

Compare readers

After a successful extraction, choose another reading engine and select Test same inputs with selected reader. This reuses the last successful test's brief, conversation and template inputs in the current session. It uses the currently saved document properties; changing those properties also changes the test. Use Run extraction / apply feedback instead when you want to test revised instructions or labels.

Under Results, expand Compare structured tests and choose two retained tests. Each shows its original reader/model, brief and read-only JSON/CSV output. Enlarge the workspace for side-by-side columns. AgentData reports whether the source version and recorded extraction inputs match; a warning means the differences cannot be attributed only to the reader. Older results without comparison metadata also show this warning. Results remain private to your account and subject to retention; expired results cannot be restored through comparison.

Process a document: brief, labels, discussion, results and history

Process document is the single entry point for scanning and AI extraction. It opens the saved workspace without starting OCR or an AI call.

1. Brief: choose one Extraction template: a built-in document type or your saved template. The template supplies its extraction behavior; there is no separate skill selector. Enter what to extract, and review language, quality, handwriting, tables and image settings. Document details is collapsed initially: open it to edit the draft name and description and view discovery/scan dates. Built-in template choices supply a default name and description. This draft name does not rename the source file. The red Save button appears at the top when changes are pending and saves both the brief and document properties. Next: optional labels is beside it and saves before advancing. Nothing is processed here. Loading a saved template starts a discussion for this document; it does not copy another document's chat.
2. Labels (optional): draw named fields or one example repeating row. Draft labels autosave; choose Save labels and draft for an explicit save. The enlarged and separate-window views show the original beside the label controls. Choose Save and continue / skip labels to save before step 3; drawing labels is optional.
3. Test and compare: select an available reading engine and, for Local or Docling, Automatic or Always OCR recognition. The active brief is visible above the previous messages. Use Edit brief to update the goal, then refine the request in chat. Run extraction / apply feedback reuses compatible retained text when available and produces a trial JSON/table result. Rescan and test starts a new recognition run before testing the current prompt. Progress and completion appear in the workspace. Generate prompt consolidates the conversation without testing extraction; Save prompt to brief replaces the instructions in the first section. You can return to Brief and edit them.
4. Scan history: the table shows start time, status, elapsed duration, reader/model, input/output tokens and estimated USD. Duration for queued jobs includes waiting time. Choose Open result to compare a retained result on the left with the original document on the right; the run table collapses to make room and can be reopened to choose another run. Processing log is here, collapsed initially. Local and Docling reading have no LLM token charge, while subsequent AI extraction is recorded separately. Costs use configured rates and are estimates, not invoices or infrastructure costs. Known usage only; total unknown and a ≥ amount mean some tokens were recorded but other calls have unknown usage. Usage not recorded does not mean free. Older missing tokens cannot be reconstructed.
5. Results: inspect expandable JSON or CSV / Table separately from the conversation. OCR text is a diagnostic input, not a structured result. Rename output columns to update the same template mapping. Enter a template name here and choose Save template, or Save and continue to extraction for the review workflow. Saving a template creates an immutable version; trial values still require extraction and human review before publication.

You can browse Brief, Labels or Scan history while a test runs and select the reader for the next test. The running test keeps the reader and settings it started with; another extraction waits until it finishes. If the response takes too long, Stop waiting releases the browser controls. Processing may still continue on the server: inspect Scan history before starting another test. Completed structured results remain visible when document metadata refreshes.

Brief, label and discussion changes autosave privately for your account in this document's workspace. Save also preserves unsent composer text. The saved draft follows the connected folder and file path across inventory refreshes. Concurrent edits in another window produce a conflict instead of overwriting the saved discussion. Wait for autosave or an explicit save confirmation before reloading; edits made during a failed or unfinished save can still be lost. Saving a reusable template separately includes its instructions and conversation in template version history and, when configured, its repository history.

Use Expand output preview or Open output in a new window at the top of the workspace for a larger view: configuration, discussion and history stay on the left, with the original document on the right during Brief, Labels and Scan history, and results on the right during Test and compare. Drag the expanded window by its title bar and resize it from the bottom-right corner to compare results with the original. The move and resize handles also accept arrow keys. The window does not block uncovered parts of the original viewer; the reset button restores its initial size and position. Small windows and screens stack the panels. Continue chatting in either large view; returning preserves the current conversation, unsent text, mapping and result. Close the separate window or choose Return results and chat to this document to return. Closing the document workspace closes its separate window too.

Diagnostics: reader text (not the extracted result) is collapsed by default and shows retained pages. Expand a page and use Load more pages for longer documents. Mixed PDFs show OCR text first and embedded text separately. Missing text and expired results display guidance rather than empty output. Open Scan history → Processing log for recognition and model details. Extraction issues beside the result explain failed model responses or incomplete coverage.

Template mapping shows the full definition; Extraction details and evidence shows trial fields, evidence and review states. Prompt or mapping changes require another Run extraction / apply feedback to refresh trial values. Before a trial runs, JSON/CSV shows the requested layout with empty values and explains that extraction has not run. Choose Extract structured JSON / CSV or Run extraction / apply feedback to create values. Saving a brief-only template does not lock it into one text column: the first prompt/test can build separate fields and repeated rows from the goal. Existing designed field types and row assignments are preserved. If an older draft already has the wrong columns or one generic text field, return to Brief → Rebuild layout from brief. After confirmation, this saves an empty field mapping for the current draft while keeping the brief and discussion. It clears draft fields, labels, row groups, derived mappings and validation rules; saved template versions and earlier results remain available. Continue to Test and compare → Run extraction / apply feedback to generate the new structure. After a trial, the result reports extracted row and value counts; JSON row contents open by default. Previous trials are read-only and remain separate from the current editable mapping.

Chat and prompt generation require processing consent and a configured AI provider. Large documents may use a text sample for prompt drafting; trial extraction processes the retained parse within configured limits and reports incomplete coverage. Trial content expires under the processing retention policy; operational history remains, and saved briefs/discussions remain user-authored configuration. Use Review → Extraction results for durable extraction and approval.

Example: blood-test rows

Choose General document, then use Blood test example in an empty chat or enter:

Extract one row per blood test: test name, result exactly as printed (including < or >), units, reference range, remarks and test date. Preserve Hebrew names and remarks. Leave missing or unclear values empty and flag them for review. Do not infer a date or clinical meaning.

Choose Run extraction / apply feedback, inspect CSV / Table, and refine with instructions such as “Keep the Hebrew remarks” or “Add the date printed in the report header to each test row.” Inspect the original and source evidence before accepting values, especially signs, decimal separators, units and dates. This extracts reported data; it does not interpret the clinical significance of results.

Contract and policy skills help structure claims, obligations, remedies, limits and conditions. They do not determine legal entitlement or approve a claim. Check the original and evidence before relying on a result.

Choose a document type

Under Templates → Start with a document type, choose a starting draft and review its fields before saving:

Starting template	What it extracts
Invoice	Invoice number, issue date, currency and total, plus each billed item and amount.
Contract	Parties, dates and repeated sections containing obligations, measures, thresholds and units.
Insurance policy	Policy identifier, insured party, effective dates, currency and repeated items with limits.

Each draft provides different instructions and row definitions to the extraction model. Customize field definitions, required fields, number/date conventions and repeated groups for your documents. These drafts are starting points, not trained or certified templates.

Extracting a contractual threshold does not establish a violation: that requires evidence of actual performance and explicit comparison rules. Claim-to-policy comparison likewise requires matching identifiers, periods, currencies and items, plus exclusions and coverage rules. Automatic cross-template matching and coverage decisions are not currently available in Document Automation.

Label examples and evaluate extraction

1. Parse the document and save a template first. If parsing is disabled, ask your tenant administrator to enable processing in Documents Automation settings.
2. Open Review → Reviewed labels, then choose the template and a successfully or partially parsed document.
3. Select the field to label. Click a highlighted word on the page image, then choose Attach selected evidence. The first selection fills the expected value; check and correct it. Attach additional words when needed, or select an exact passage under Extracted text.
4. For repeating invoice items, contract sections, or policy items, add a row with a unique row key and label its fields separately.
5. Enter a review reason and choose Approve and save labels. Only mark a field absent after complete parsing. If OCR missed the text, obtain a clearer original and reprocess; template rectangles cannot substitute for verified text evidence.

Suggestions are editable drafts; review them before approval. Keep calibration examples separate from unseen holdout documents.

The initial pilot uses 20 historical documents, including an unseen holdout subset. Measure field values, required coverage, evidence and repeated rows against approved labels. Snapshot agreement or a successful job alone is not an extraction-accuracy certificate.

Extract and review results

1. Ask your tenant administrator to enable processing consent and retention in Settings & Monitoring → Settings → Documents Automation. Preview/extraction may transfer document contents to the application worker; the folder connector does not imply processing remains on your computer.
2. In the document chat panel, choose Save and continue to extraction. The reviewed document parse and saved template version are selected for you. Alternatively, open Review → Extraction results and select them yourself. Supply any declared runtime inputs, then choose Extract fields. A progress indicator runs until extraction completes, and the completed result opens for review.
3. Open Review result. Check required/unresolved counts, page coverage, source evidence, missing rows and conflicts. Open source page highlights the referenced evidence.
4. Enter a reason when correcting, accepting or rejecting a result. Corrections and decisions create new revisions. Blocking or incomplete output cannot be accepted.
5. Export accepted tables or download the JSON/CSV representations. Review states and source/result versions remain explicit.

Human acceptance is required. Changing templates, labels or execution configuration can make prior evaluation evidence stale; rerun affected checks.

Monitor and schedule folders

Open Connectors in the workspace navigation, then expand Show folders and scan settings. Each row has one folder policy:

• Manual: choose Scan to refresh the inventory, then Process document for an individual file.
• On change: AgentData checks for new or changed versions at the interval shown in the section (normally every five minutes) and queues their OCR/text processing.
• Scheduled: enter the interval in minutes, then Save. The existing AgentData scheduler scans and processes new or changed files at that interval.

Automatic policies require processing consent and a running customer-side connector. They use periodic checks, not instant operating-system notifications. Unchanged known versions are deduplicated. Saving a folder policy does not automatically extract fields or accept results.

For recurring extraction with a specific template, open Review → Extraction results, select a template/version and runtime inputs, choose Scheduled folder and an interval, then Schedule folder. These extraction schedules remain independent of the folder's scan policy; the folder row indicates when additional extraction schedules exist.

Monitoring shows processing jobs and scan schedules, including failures and next-run times. Disable a schedule to stop its pending work. Consent or access changes can halt processing; after resolving the cause, save the folder policy again or recreate the extraction schedule. Review contains extracted results and their publication status.

Accepted outputs

Extracted values and review revisions are saved in the AgentData database. With tenant Git synchronization configured, new revisions are also committed under documents/projects/<project-id>/<document-id>/runs/<run-id>/revisions/<revision>/result.json inside the configured repository directory. Results show the Git synchronization state; a Git error does not remove the saved database result. Original files stay in their source folder. Download accepted JSON/CSV from Results; local or S3 folder-write permission is unnecessary. Git snapshots contain extracted values and review metadata, and follow the repository’s own history/retention.

Publish reviewed document tables for SQL

1. Open an accepted result in Review. Under Publication, choose an authorized direct PostgreSQL destination.
2. Choose Read connection for SQL queries for the same database. If none is listed, add the PostgreSQL connection in Connectors, assign it to this project, and refresh the destinations. Its database account must be able to read the published tables. Write and read accounts may have different permissions.
3. Enter a destination table name for each output table and choose Publish accepted result. AgentData adds a tenant/project prefix to the physical names and preserves the declared column types and row keys.
4. Wait for a successful receipt, then choose Open … in Datasets / SQL. Publication checks read access before reporting the dataset as ready. If registration fails after the database write, fix the reported issue and retry the publication; retrying does not duplicate the rows.

A later reviewed run for the same source path, template version and runtime inputs replaces that document's previously published rows, including removed child rows. Other documents remain intact. An older run cannot overwrite a newer published run. Changing a template version or table schema requires a new destination table name; existing tables are not silently migrated.

Compare document data with relational data

Open Datasets / SQL, select a published dataset, then optionally choose a confirmed relational table on the same read connection. Discover and confirm the reference table in the semantic model first if it is missing.

Choose explicit matching keys—for example, a test code and unit, or a policy item, currency and period. Add multiple keys where needed. Keep unmatched document rows uses a left join; unmatched reference values appear as NULL. Matching rows only excludes them. Names, currencies and units are not matched or converted automatically.

Choose All rows to inspect the joined values, or select a comparison and its document/reference columns. Run SQL query displays the rows and lets you expand the generated SQL and parameters. Queries are read-only and bounded; a message appears if the row limit is reached. Results are from the last query; rerun after changing the controls.

Editors may save the matching keys in the semantic model. Choose the relationship cardinality; AgentData checks uniqueness on the declared “one” side before saving. Published tables are registered as semantic entities, and their Cube/dbt YAML and saved relationships use the tenant's configured Git synchronization. Future changes to reference data may change key uniqueness, so maintain the reference database's constraints as well.

This screen supports comparisons within one PostgreSQL read connection. Cross-database semantic joins require the separately configured federation capability. Only reviewed, published rows are queried: unpublished documents and unresolved extraction results are not part of the dataset. A query result does not certify complete folder coverage.

Keep numeric qualifiers and original text

For values such as <0.5, keep the original field as text. In the extraction preview's CSV / Table view, choose Add numeric columns for that text field. AgentData adds a numeric value and operator derived from the same source evidence. Run Run extraction / apply feedback again, then save and extract using the updated template.

For example, <0.5 becomes original text <0.5, numeric bound 0.5, and operator <. An unqualified 92 produces 92 and =. Missing values remain missing. Ambiguous ranges, prose or undeclared number formats require review rather than a guessed number. Set an explicit numeric locale in the template schema for locale-dependent separators.

Numeric comparison filters exclude qualified bounds and missing values; All rows retains them for inspection. No medical, legal or financial interpretation is inferred. Chat revisions preserve existing types, required fields, validation rules, runtime inputs and output mappings. Remove fields or change their types/groups in the template schema editor before asking chat to revise the instructions.

Automatic Flow events and automatic acceptance remain unavailable.

Troubleshooting

State or issue	What to do
Processing disabled	Ask a tenant administrator to review and enable processing consent.
Not scanned / empty inventory	Confirm the folder contains supported documents, then scan.
Partial scan	Resolve the reported access or inventory-limit issue before relying on the count.
Connector offline	Start the connector and check its local configuration, then rescan.
Unsupported / blocked file	Check its format and path. Safe document reads reject symlinks and special files.
Original changed since scan	Rescan before opening or processing the new version.
Queue or transfer limit	Wait for existing work or ask the administrator to review configured limits.
DOC-PARSER-FAILED / parser memory limit	Choose Process document → Test and compare → Rescan and test after the issue is resolved. Check the processing log; if it repeats, share the error code with your administrator. A failed job does not certify or accept output.
Incomplete parsing / OCR	Review page coverage and issues; reprocess after the issue is resolved.
Expired artifact	Reprocess the authorized original; an expired retained artifact cannot be recovered through its old result.
No recognized text / rotated or unclear scan	Correct orientation or supply a clearer scan, reprocess, and review numeric values against the image. A completed parse does not certify OCR accuracy.

Access is scoped to tenant/project and rechecked when reading originals, results and exports. Full enterprise delegated document access, Windows document-folder access and customer/provider certification remain separate availability gates. See Enterprise identity for supported identity behavior.

For versioned App and Data Flow authoring, test results and activation, see Apps and Data Flow Notebooks.

If a document preview fails

Choose Retry preview after checking that your folder connector is running. AgentData shows the reason when the source changed or a rendering limit was reached. For a changed file, open Connectors, scan the folder, then reopen the document. Download original is available for viewing the source locally. A failed preview is not marked as processed, and retries do not discard your unsaved template labels.

Preview speed and document storage

The first view reads the original from its source and renders a page image. A local folder connector must be online; transfer time and large or complex PDFs can affect loading. Recently rendered pages are reused briefly when the source bytes are unchanged. AgentData still checks access and reads the current source, so reopening a preview can still involve a connector round trip.

Local-folder originals remain on your machine. Managed uploads and retained original copies used for extraction are encrypted in AgentData’s database and subject to retention limits. Structured extraction results are also stored in the database. Simply viewing a local file does not permanently import its original into the database. More storage is only needed when capacity or quota is the problem; it does not by itself speed up connector transfers or PDF rendering.