feat(examples): vision-first document RAG (ColQwen2.5 + Florence-2-DocVQA)

[svonava]

Summary

• New example under examples/vision-doc-rag/. ColQwen2.5 ranks pages by reading them as images, Florence-2-FT-DocVQA reads the top page and produces a textual answer. OCR never enters the score path, so charts, tables, screenshots, and layout cues survive end-to-end.
• Multi-tenant from the start: every page carries a client tag, queries are scoped via a Python filter before MaxSim. Same corpus serves multiple tenants with no per-tenant index.
• Optional Qwen/Qwen3-VL-Reranker-2B second stage stays in the visual modality (off by default — gated on a cluster-side bugfix).
• Self-contained: 12 synthetic pages across 3 fictional clients, a PIL renderer that turns each entry into a PNG, FastAPI server, minimal UI that shows the page image alongside the answer.

SIE features

Stage	Model	Primitive
Retrieval	vidore/colqwen2.5-v0.2	encode (multivector, image + text)
Ranking	client-side	sie_sdk.scoring.maxsim
Rerank (optional)	Qwen/Qwen3-VL-Reranker-2B	score
Answer	mynkchaudhry/Florence-2-FT-DocVQA	extract with instruction=<question>
OCR snippet (UI only)	mynkchaudhry/Florence-2-FT-DocVQA	extract

Project layout

examples/vision-doc-rag/
├── README.md
├── config.yaml
├── data/
│   ├── fetch_dataset.py    # synthetic 3-tenant corpus → pages.json
│   └── render_pages.py     # pages.json → PNG screenshots
├── python/
│   ├── ingest.py           # encode every page → multivectors.npz
│   ├── search.py           # CLI demo: 4 scoped queries with timings
│   ├── server.py           # FastAPI /api/search?q=&client=
│   └── requirements.txt
└── static/
    └── index.html          # tenant selector + query box + answer card

Test plan

• data/fetch_dataset.py generates 12 pages across 3 tenants
• data/render_pages.py renders 12 PNGs (1024×1280) via PIL with a fallback font path
• First page encode via vidore/colqwen2.5-v0.2 returns a [~740, 128] multivector on the dev cluster (verified before a cluster-side wedge took the worker out — see notes)
• sie_sdk.scoring.maxsim(query_mv, [doc_mv]) returns the expected high score for a matched page
• BAAI/bge-reranker-v2-m3 returns sensible scores on text inputs (sanity test, separate from the visual reranker)
• End-to-end ingest + search against a healthy SIE cluster (blocked on the cluster recovery I've been working through with @ - internal context)
• DocVQA instruction=<question> returns a focused answer rather than an OCR dump (depends on the cluster's Florence-2 adapter routing the task token correctly)

Notes for reviewers

The visual reranker is configured but disabled by default in config.yaml. There's a known cluster-side adapter issue where JSON image inputs are not base64-decoded before reaching the preprocessor; once that lands, flip search.visual_rerank: true and the second stage runs in the same modality as retrieval.

The synthetic corpus is intentionally domain-mixed (engineering runbooks, HR policies, finance procedures) so queries clearly disambiguate by tenant and the visual layout matters more than keyword overlap.