Document Insights

Explore documents through extraction profiles, evidence-linked fields, tables, signatures, terms, and chat.

Document Insights combines managed file storage, extraction workflows, a similarity index, source viewing, and document chat. Its catalogue identifier is pdf-insights. The interface helps you move between an original document and structured results instead of relying only on a generated summary.

Upload and ingest

Create Document Insights and use Drop documents to process to upload and ingest documents. To bring in an existing provider file, use Add to App files or drag it into the processing area. Uploading into a provider folder alone does not start this ingestion flow. Follow run status, then select a processed document. The ingestion workflow prepares source content; extraction and chat are separate tasks that use that material. A stored file can exist before extraction results are ready.

Choose an extraction profile

Select a profile that matches the material, then choose Extract. The template offers auto-discovery, invoice, employment or contractor agreement, generic document, and investors’ rights agreement choices. Auto-discovery uses an additional model call to propose an appropriate shape. It is useful when you do not know the document’s structure, but a selected profile gives a more deliberate field set.

The profile and extraction-budget controls affect the extraction task. Review the budget before processing long or numerous documents. Changing a profile does not retroactively change an earlier extraction: inspect which run and profile produced the selected result.

Inspect the results

The field table shows field, value, evidence, and page. Select a discovered table to see its rows. Other views expose signatures, material terms, and issues or time windows where returned by extraction. Signature rows can include page, entity, signatory, title, and status. Terms include the term label, value, evidence, and page.

Use the document viewer to check each important value against its source. A page number or quoted span helps locate evidence; it does not guarantee that the extracted interpretation is complete. A signature-like mark is not itself a decision that an agreement is effective. Missing fields can mean the information is absent, not found, or incompatible with the selected profile.

Loaded results and full run counts

The starting template reads a sequence-ordered window of at most 200 extraction records. A Showing N of M notice means the run produced more records than the current results read loaded. Fields, table rows, signatures, or terms beyond that window may therefore be absent from the interface even after a successful extraction. Check the full run summary and inspect or expand the results read deliberately; rerunning extraction does not remove the display cap.

Ask questions

Use chat for questions about the ingested documents. Give a specific question and identify the document when the app holds several. Open supporting material before treating the answer as a verified fact. A relevant retrieved passage is different from proof that the answer considered every page.

Refine and recover

Try a more suitable profile or improved source if the structure is wrong. Inspect saved outputs before repeating an uncertain extraction. Use document profiles, tables, and workflow run monitoring for deeper controls. Exporting the app definition preserves its design separately from live files, rows, and extraction results.

Parameter reference

These are the starting template’s declared parameters. Your saved app can have its own version of the schema. A default applies when no explicit value is supplied; a required field still needs a valid value.

• title: exposed as an app interface parameter.

Connection slots

Choose compatible saved connections for user-supplied slots. Managed slots are supplied by Strive. Disabling an optional slot stops its use without necessarily deleting the saved binding. See connector policy.

Included workflows

Inspect the saved app’s workflows for changes made after creation. A workflow’s presence does not mean that every interaction runs every workflow.

Output format: extraction_result

One result of a document extraction, discriminated by kind: the run summary (counts, scores, tasks), a field or term with its page, quote, verification status and error_code, a table row (its table's columns beside table), a signature, a quality check (rule, score, passed), an issue (task, code, message) or a model window. seq is its display slot.

Output format: message

One assistant reply in the document chat: the question it answered (input.message), the passages vector search retrieved for it (matches) and the model's answer (content).

Output format: message_run_marker

Marks the chat run whose answer the App shows. Written by the mark step as the run's completion marker.