Add azure-di-financial-haystack integration

[zavera]

Adds an integration listing for azure-di-financial-haystack — structured KV extraction and delta reconciliation for financial PDFs using Azure Document Intelligence.

Components:

• AzureDiExtractor — 4-stage PDF recovery chain, multi-endpoint round-robin pool
• KvNormalizer — maps verbose Azure DI labels to canonical field names, handles punctuation/whitespace/newline variants
• DeltaCalculator — compares extracted values against reference values, scores HIGH/MEDIUM/LOW
• build_pipeline() — pre-wired convenience pipeline

Tests: 76 unit tests passing across Python 3.10–3.13, no Azure credentials required. Live Azure integration tests available under pytest -m integration.

PyPI: https://pypi.org/project/azure-di-financial-haystack
Repo: https://github.com/zavera/haystack-financial-doc-extractor

[vercel]

@azaver1 is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[zavera]

The package is now live on PyPI: https://pypi.org/project/azure-di-financial-haystack/ (v0.1.1). Also fixed the LinkedIn URL in the latest commit. Please take another look when you get a chance!

[zavera]

Hi @kacperlukawski just following up on this. The LinkedIn URL fix is in and the package is live on PyPI at v0.1.1. Happy to address any remaining feedback whenever you get a chance to take another look.

[zavera]

Quick update on azure-di-financial-haystack, now at v0.3.0 on PyPI. This package is Azure + IRS specialized:

• Language support: AzureDiExtractor detects each document's language from Azure DI's own result and translates non-English documents to English via Azure OpenAI before extraction.
• IRS form classification: new IrsFormClassifier, an LLM reasoning component with a purpose-built prompt that identifies every IRS form type in a document, including bundled multi-form uploads (e.g. Schedule C plus Schedule SE in one file).
• get_content() / get_kv_pairs(): read a document's full extracted text or structured KV pairs off a result. get_content() feeds the classification prompt above.

Original PyPI-link concern should be resolved (live since v0.1.1). Happy to answer questions, and a re-review when convenient would help.