Latest in AI
Amazon Textract Insights: Review a Mixed PDF Before Importing Its Fields
Build a mixed-packet extraction pilot that checks document splits, blueprint matches and source evidence before writing to downstream systems.
A successful document extraction job can still leave an invoice unclassified or put a delivery note into the wrong record. With Amazon Textract Insights, the status worth checking is not just whether the API finished. It's whether every part of the packet made it through the right blueprint and whether the fields you care about have usable evidence.
AWS made Textract Insights generally available on November 20. It is a separate API workflow from Textract Core. For a first pilot, I would run it beside an existing importer rather than connect it directly to production records.
Pick a packet with a known answer
Start with a sample supplier packet containing a two-page invoice, a delivery receipt and a cover note. Before submitting anything, write down the expected page groups and which fields belong to each. Use synthetic or approved test documents.
A blueprint describes the fields to extract. The custom extraction guide lists public templates, including an invoice blueprint, and explains how to add your own. Start with an invoice blueprint and a narrow delivery-receipt blueprint. Leave the cover note outside both so the pilot can exercise the unmatched path.
Define the fields precisely. For example, identify the supplier's invoice number separately from your purchase order number. If a field is supposed to match text printed on the page, don't ask the service to infer it from related text. Keep any inferred fields separate so reviewers know which values call for judgment.
Submit one job and read the manifest
The AWS first-document tutorial shows the asynchronous sequence: create a project, submit an S3 input through StartDocumentInsights, then use the returned job ARN with GetDocumentInsights. For a mixed packet, enable splitting on the project before submitting it.
Use a dedicated S3 output prefix for the pilot. Once the job succeeds, read job_metadata.json before reading individual result files. Follow the paths listed in its segment records rather than constructing a path for each expected document. A run that produces two segments when you expected three should not quietly create two clean-looking records.
Check what processing boundary your documents require before the pilot. The launch announcement distinguishes Insights, which can route inference across Regions within the originating geography, from Core, which processes requests in the called Region. Don't infer the Insights boundary from a Core deployment.
Handle unmatched documents even when the job succeeds
The blueprint matching documentation distinguishes MATCH, NO_MATCH and FALLBACK at the segment level. A NO_MATCH segment can leave the overall job in SUCCEEDED and still have no custom result path. One primary blueprint without a fallback skips matching altogether, so a match confidence of 1 in that setup doesn't establish the document type.
For the supplier packet, review the invoice and receipt as separate records and flag the cover note as unclassified. If you add a fallback later, send those results to review first. A fallback is useful for surfacing information, but treating it as a confirmed document class would hide the distinction the status field was designed to preserve.
Check grounding before setting an acceptance threshold
In the custom result, the tutorial shows values under inference_result and confidence information in the corresponding explainability_info. AWS describes grounding that points reviewers to source pages and page coordinates.
Give reviewers the proposed value next to its source. For the invoice total, check that the evidence shows the final amount rather than a subtotal. For the receipt date, check that the cited text belongs to the receipt rather than the adjacent invoice. A confidence score cannot settle either question on its own.
Track separate errors for wrong splits, wrong blueprint matches, missing values and incorrect values. Compare those results across clean and degraded scans before choosing field-specific review rules. That will tell you whether to fix the document grouping, revise a blueprint or keep a human checkpoint.
Keep the first run read-only. A useful pilot produces a record you can trace back to the packet and a review queue that makes missing or uncertain results visible. Only then decide what is safe to import automatically.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams explain technical products through clear documentation, practical guides and developer-focused content.