Changelog

Table cells you can trust, and pipeline retries. Full write-up in the release post.

• Cell-level table provenance (bounding box + source page per cell); row and column spans preserved instead of splitting into ghost rows.
• Per-page pipeline retries — partial results survive, failed ranges are retried, webhooks carry a partial flag.
• heredoc schemas in YAML; unstructred diff CLI for comparing pipeline outputs across versions.
• 30% latency cut for TIFF-heavy workloads. Dropped Python 3.8 support (it had been EOL for over a year; we were sentimental, then we weren't).

• AsyncClient out of beta, same API surface as the sync client.
• Result webhooks with HMAC signatures and 24-hour retry.
• Batch endpoint raised from 1,000 to 10,000 documents per submission.
• Handwriting extraction (beta) promoted from "we're trying it" to "we support it, with caveats in the docs."

• Pipelines: pipeline.yaml, watchers on S3 buckets and inboxes, sinks to webhooks and HTTP.
• Confidence thresholds per field (route anything under 0.8 to human review, keep the rest, sleep at night).
• First 300 customers hit the platform somewhere around here. We noticed because the ingest graphs went up and stayed up.

• Every extracted value carries a confidence score between 0 and 1.
• Per-field bounding boxes and page numbers in results.
• API v2 paths (/v1/parse, /v1/batches, …) — yes, the SDK is v2 and the API paths are v1; naming is hard and we live with it.

• Fax decoding (T.38 and T.30 scans). Healthcare customers rejoiced; the rest of you may be surprised faxes still exist. They do. Oh, they do.
• Handwriting extraction, private beta — Portuguese pharmacy labels, German forms, one memorable batch of Icelandic fishing logs.

• TIFF support, multi-page and single-strip.
• Table extraction for line items (no provenance yet — that's the v2.3.0 story).
• Fixed a parser panic on PDFs with circular xref tables. Adobe Reader opened them fine, which we found insulting.

• .eml and .msg email parsing with attachment handling.
• CSV export of parse results.
• Sandbox environment with fixture documents and test keys.

• Public launch: PDF parsing, schema-driven extraction, the Python SDK.
• The company was founded in June 2024; v1.0 shipped 105 days later, which either sounds fast or slow depending on how many parsers you've written.