feat(celery Wave 6 #39): per-provider multimodal embedding input payload + fix self.base_url bug

[earayu]

Summary

Wave 6 task #39 — per-provider multimodal embedding input payload dispatcher + latent Wave 5 P2 chunk 1 bug fix.

Wave 5 P2 chunk 1 shipped EmbeddingService._embed_image_via_litellm with a single LiteLLM-shaped payload that mirrored OpenAI's multimodal chat completion request (not the embedding request). Voyage / Jina / Cohere each document a different embedding input shape; LiteLLM may translate transparently for some but ships no guarantee. This PR adds aperag/llm/embed/multimodal_input.py:build_multimodal_input_payload to emit the canonical shape per provider, with the Wave 5 baseline as the unknown-provider fallback.

Also fixes self.base_url → self.api_base in the same call (constructor sets self.api_base = embedding_service_url; the existing reference would have raised AttributeError on the first real production embed_image call).

Per-provider shapes

Provider	Aliases	Shape
Voyage AI	voyage_ai / voyageai / voyage	[{"content": [{"type":"image_base64","image_base64":<data url>}, {"type":"text","text":<alt>}]}]
Jina	jina_ai / jinaai / jina	[{"image":<data url>}, {"text":<alt>}] (fused vector)
Cohere	cohere	[{"image":<data url>}, {"text":<alt>}]
OpenAI	openai / openai_multimodal	[{"type":"image_url","image_url":{"url":<data url>}}, {"type":"text","text":<alt>}]
Unknown	(fallback)	[{"image_url":{"url":<data url>}}, {"text":<alt>}] (Wave 5 P2 baseline preserved)

Empty / whitespace alt_text skips the text part on every provider — pairing the image with " " would change the cache key (#37) and potentially confuse the embedder.

Bug fix

embedding_service.py:_embed_image_via_litellm referenced self.base_url (undefined) in the LiteLLM call; should be self.api_base. Verified via grep: zero other self.base_url references in this file. The text path at line 323 already uses self.api_base correctly. Without this fix, the first production embed_image() call raises AttributeError: 'EmbeddingService' object has no attribute 'base_url'.

Tests

tests/unit_test/llm/test_multimodal_input.py — 11 tests:

• ✅ Voyage content envelope + image_base64 part + alias matching (voyage / voyageai / VOYAGE / Voyage)
• ✅ Voyage / Jina / Cohere / OpenAI / unknown empty-alt_text path skips text part
• ✅ Jina + Cohere flat-list shape
• ✅ OpenAI chat-multimodal envelope + alias openai_multimodal resolves same shape
• ✅ Unknown provider falls back to LiteLLM-baseline shape
• ✅ provider=None resolves to default
• ✅ Whitespace-only alt_text treated as empty across all providers (regression guard)

Full unit suite: 1038 passed, 29 skipped (16 new + 1022 from main), ruff + format clean.

Out of scope

• No new operator config — provider keyword already drives the rest of the embedding stack (cache key feat: async running #37, text path, error wrap).
• No backend / column rename — task chore: change apecd #36 territory.
• No cache layer changes — task feat: async running #37 wiring already shipped.

Simple-stable 4 guardrail check

Guardrail	#39 verdict
不无限扩范围	✅ scope = single dispatcher module + 1-line callsite swap + 1-char bug fix
先把功能做实	✅ each provider gets the documented embedding wire shape; latent bug fixed so production path is reachable
简单稳定	✅ pure function dispatcher, string-match by lowercase, no abstraction layer
私有化免维护	✅ provider keyword already operator-set; no new knob

Test plan

• uv run pytest tests/unit_test/ — 1038 passed
• uvx ruff check + format — clean
• base on 0edc82a6 (post-Wave 5 + feat(celery Wave 6 #37): wire EmbeddingService.embed_image into application cache #1735 + docs(celery Wave 6 §K.11): full spec amendment — 7 items + 2-PR strategy + double pre-check lock #1736 merged main)