Field Notes

Reliably read documents, answer tables, build workflows: three strong updates

You get auditable fields from PDFs, answers from Excel, and repeatable team routines with no code.

The week’s key AI news

  • Amazon Textract Insights is generally available. You describe fields in natural language and get structured values back with a confidence score and page reference. Mixed PDF bundles with up to 3,000 pages are logically split and mapped to matching templates. The basis is so-called Blueprints, a JSON schema with field descriptions and an inference type per field. Blueprint: a structured field list as a JSON Schema. There are public Blueprints for W-2, 1040, 1098, passport, and more, and you can bring your own. Textract Core also improved at reading, including dense tables, handwriting, and 40 languages in DetectDocumentText and AnalyzeDocument.1
  • Amazon Bedrock Managed Knowledge Base now does table search. Questions that need data from CSV or Excel run as code in an isolated sandbox over the whole table, not just a few retrieved passages. The agent links computed values from tables with facts from documents and cites both. The trace names the workbook it read. Table search: the agent reads the whole table and computes over it. It works through existing connectors like S3, SharePoint, or Confluence and uses the same AgenticRetrieveStream request. Answers take longer than text-only and do not replace BI reports, but they close gaps on thresholds, sums, and lookups.2
  • AWS Quick gets new Flows, an updated web app, and multi-agent orchestration. You describe processes in chat, Quick builds a flow with deterministic steps, and it can handle longer-running workflows with large data volumes. The web app shows a daily briefing, keeps work moving, and brings skills to the web. New sources: Meta Ads and Amplitude. Plus configurable safety and sensitive controls and language support including German. The stated time savings of a customer come from vendor claims.3
  • For self-hosted LLMs, AWS presents an agent skill for distributed KV caching. Shared KV caches avoid re-prefills of long contexts after routing switches. AWS says up to 40 percent cost savings and 80 percent faster time-to-first-token are possible. The skill guides you through suitability check, size estimation with KVTrace, setup with vLLM or SGLang, and trial runs. Watch tenant isolation when response times could leak signals. The numbers come from the vendor.4
  • Standard Bank automates risk review for GenAI use cases. Intake via the Well-Architected Tool, orchestration via Step Functions, evaluation with an LLM judge on Bedrock. The pipeline run checks documentation, code, and live calls, adds jailbreak prompts, and tests meaning consistency across multiple paraphrases. Everything runs in your own AWS account. Results land in a standard format for the committee. Good if your approval process slows things down.5
  • Apify launches Server Actors. Agents get real-time answers with low latency and the platform scales servers dynamically. For large batches the classic input mode stays cheaper. The note comes from a partner post.6
  • The trend toward formal verification is picking up speed. Vendors like Harmonic, Theorem, Axiom Math, and &Truth combine AI with proof assistants like Lean. A practical focus is memory safety checking for C code, with Lean verifying every step. For teams with lots of generated code this can cut slop risk.7
  • Mathematician Levent Alpöge writes that he found a proof of the Riemann Hypothesis with Claude and formalized it in Lean. That is a claim, not a verified fact. Nothing follows for daily work for now.8

The week for your company

  1. Reliably read mixed PDF bundles. If you have mortgage files, claims, or onboarding packets, test Textract Insights with a public Blueprint plus one or two custom fields. You get values with a confidence score and page coordinates and can focus reviews on low scores. You need S3 and access to Textract.1
  2. Ask questions of tables instead of calculating by hand. Create a Knowledge Base with your final Excel and the right policy and ask a question that needs both. Name the file in the prompt so retrieval finds the right table. Check citations and trace. Plan a bit more runtime than for text-only.2
  3. Capture a recurring routine as a Quick Flow. Describe the process in chat, add fixed calculation steps, and let multiple agents work in parallel. Turn on the new safety controls and share the flow with the team. The monthly report keeps running even when you are on the road.3
  4. Speed up self-hosted LLMs. Use the KV caching skill to estimate benefit with KVTrace, then stand up a test cache layer. Keep KV IDs and data types consistent across all inference servers, or you get hit counters without correct answers. Plan separate caches for strictly isolated tenants.4
  5. Speed up governance instead of stopping it. Follow the Standard Bank pattern. Automated evidence collection, defined controls, live tests with red-team prompts, and standardized reports. The committee still decides, but on a consistent data base. Start with one pilot use case.5

Try next week

  1. Set up a small collection in Bedrock Managed Knowledge Base: one Excel or CSV with numbers plus two short PDFs with the related rule and a memo, via S3 or SharePoint.2
  2. Ask a question with a calculation and name the workbook in the prompt. Open citations and trace and check if the agent read the right file and applied the rule correctly.2
  3. Note the runtime. Mark questions that belong in BI because they recur or combine multiple sources.2

A good find

  • Linear’s COO shares three hard learnings: measure agent usage and app usage separately. If only the agent line climbs, you become a swappable database.9
  • Support stays human-owned. AI may draft, but a person decides what goes out. This lowers the risk of embarrassing answers.9
  • Many budgets are annual. Plan for renewals instead of relying on multi-year contracts. Build measurable results in early.9

Sources

  1. Announcing Amazon Textract Insights: Verifiable generative AI for documents (aws.amazon.com)
  2. Get answers from tables with tabular search in Amazon Bedrock Managed Knowledge Base (aws.amazon.com)
  3. New in AWS Quick: More capable flows, an updated web experience, and agents that work together (aws.amazon.com)
  4. Cut self-hosted LLM cost and latency with the distributed KV caching agent skill (aws.amazon.com)
  5. How Standard Bank Group automates AI risk assessment for generative AI (aws.amazon.com)
  6. @svpino: There's only one correct way of giving your agents access to real-time web data.

I've used @apify for years now. It's the best way to use live Internet data.

Check out the list o (x.com) 7. The War Against Slop and an Emerging Weapon to Fight it: Formal Verification (latent.space) 8. Quoting Levent Alpöge (simonwillison.net) 9. Scaling Linear’s GTM to $100M ARR — Cristina Cordova, COO (latent.space)