feat(benchmark): support structured prompt tags (value+category) and origin

[RapidPoseidon]

What

Exposes structured prompt tags and a prompt origin in the Python SDK. A prompt's tags are now a list whose items are each either a plain str or a Tag(value, category=None) — you can mix both. Plain strings keep working.

SDK leg of a 3-part full-stack feature:

• Backend: RapidataAI/rapidata-backend#4879
• Frontend: RapidataAI/app-frontend#2834

Contract (as finalized by the parent task)

• High-level API is a str | Tag union: add_prompts(tags=...) / create_new_benchmark(tags=...) accept a list of str, a list of Tag, or a mix. Bare strings are wrapped to Tag(value, category=None) internally.
• Wire is clean Tag[] objects — no union type in the generated model, no taggings field, no deprecated string[]. origin is a separate additive { value } field.
• Read-back: .tags stays list[list[str]] (values, back-compat); new .structured_tags -> list[list[Tag]] for categories; new .origins. Reads defensively wrap a bare string as an uncategorized tag so an older backend still parses during rollout.

Changes

Generated client (src/rapidata/api_client/models/)

• New sub-models PromptTagging (value, optional category — the wire Tag) and PromptOrigin (value), in the generator's LazyValidatedModel style.
• tags → List[PromptTagging] and additive origin on the create input (CreatePromptForBenchmarkEndpointInput) and the read model actually returned by the prompts endpoint (GetPromptsByBenchmarkEndpointOutput).
• Hand-edited rather than regenerated: the regen script pulls specs from prod, where these fields aren't deployed yet, so a regen today would drop them. A normal regen reproduces this shape once the backend spec lands.
• ⚠️ The seed referenced PromptByBenchmarkResult, an orphan not returned by any endpoint — the live read model is GetPromptsByBenchmarkEndpointOutput. The orphan is left untouched.

High-level surface

• New user-facing dataclasses Tag(value, category=None) and Origin(value), exported from rapidata, rapidata.rapidata_client, and rapidata.types.
• add_prompts / create_new_benchmark gain the str | Tag tags union + origins; bare strings/str origins are coerced. BenchmarkPrompt carries structured tags/origin; the uploader sends clean Tag objects.
• Read accessors: .tags (values), new .structured_tags, new .origins.

Version & docs

• 3.17.2 → 3.18.0. docs/mri_advanced.md updated.

Changelog

Prompt tags now accept str | Tag — a Tag carries a value and optional category; plain strings still work and are auto-wrapped. Prompts can also carry an additive origin. Read categories via benchmark.structured_tags and origins via benchmark.origins; benchmark.tags remains the values-only view.

Deferred to the post-deploy regen pass

Two agreed-contract items — update-editable prompts and server-side filter/sort — are now finalized in backend #4879 but are not yet in the deployed spec, so they're intentionally left for the regen pass (triggered once #4879 merges & deploys) rather than hand-edited into the generated client and thrown away by that regen. Documented final shapes, captured here so nothing is lost:

• Update endpoint: prompt update accepts tags: Tag[] + origin { value } (same shape as create). Will add a high-level update_prompt(...) surface on top of the regenerated client.
• Filter (prompts listing): tag_values[in], tag_categories[in], origin_value.
• Sort: id, created_at, origin_value only — tag value/category are filter-only.

Regen-pass checklist: regenerate from the live spec → generate the backend's real Tag model → delete the prompt_tagging.py placeholder → grep -ri tagging src/ returns zero → wire the update + filter/sort high-level surface.

Validation

• python -c "from rapidata import RapidataClient; print('OK')" → OK
• pyright src/rapidata/rapidata_client → 0 errors
• black clean on changed rapidata_client files
• Round-trip + add_prompts mixed-str/Tag coercion, cache, and validation exercised locally (incl. old-backend bare-string reads)

🔗 Session: https://poseidon.rapidata.internal/chat/node-ae73dd05