fix(embedder): route batch embedding to /v1/embeddings (Issue #629)

[jlin53882]

Summary

Fix Issue #629: Ollama batch embedding fails after PR #621

Problem

PR #621 fixed Ollama 0.20.5+ single embedding by switching to /api/embeddings with prompt field. However, batch requests send payload.input as string[], causing Ollama to reject with:

"json: cannot unmarshal array into Go struct field EmbeddingRequest.prompt of type string"

Solution

Route requests by input type inside embedWithNativeFetch():

Input Type	Endpoint	Field
Single (string)	/api/embeddings	prompt (PR #621 unchanged)
Batch (string[])	/v1/embeddings	input

Additionally, add response validation to ensure batch responses contain the expected count and non-empty embeddings.

Changes

• src/embedder.ts: Route based on Array.isArray(payload.input)
• test/embedder-ollama-batch-routing.test.mjs: 5 test cases covering routing and validation

Testing

Test	Result
Single /api/embeddings with prompt	PASS
Batch /v1/embeddings with input[]	PASS
Batch count mismatch rejection	PASS
Batch empty embedding rejection	PASS
Single-element batch routing	PASS

Limitation

This assumes /v1/embeddings batch mode is valid for the target Ollama/model combination. Verified with jina-v5-retrieval-test.

Closes #629

[rwmjhb]

Clean fix for the batch embedding regression introduced by #621. The root cause diagnosis is correct and the routing split is the right approach. Ready to merge with a few minor notes.

Nice to have (non-blocking)

• Tests bind to port 11434 — the live Ollama default. Any developer running Ollama locally will hit EADDRINUSE. Consider a random available port or mock the HTTP layer instead.
• The two positive routing tests assert the response shape but not which endpoint was called — a future routing regression won't be caught. Worth asserting the request URL in at least one test.
• The new batch native-fetch branch discards provider options already computed by buildPayload() — double-check whether timeout, headers, or other options from that path are needed on the batch route.
• /v1/embeddings is used unconditionally for batch with no fallback — if a model doesn't support that endpoint, the failure will be silent from the user's perspective. A comment noting the assumption would help future maintainers.

CI note: New test file test/embedder-ollama-batch-routing.test.mjs appears unregistered in the CI test manifest — verify-ci-test-manifest.mjs may need updating before the full suite passes.

Before merge: stale_base=true — quick rebase against main to confirm nothing has touched embedder.ts there.

[jlin53882]

PR Review Updates

All review comments from @rwmjhb have been addressed:

Fixed:

1. ✅ CI manifest - Added test/embedder-ollama-batch-routing.test.mjs to scripts/ci-test-manifest.mjs
2. ✅ Port conflict - Changed test to use port 0 (OS-assigned available port) instead of hardcoded 11434, avoiding EADDRINUSE when developers have Ollama running locally
3. ✅ Endpoint verification - Test already validates which endpoint is called via capturedRoute assertion
4. ✅ Provider options - Added code comments explaining why encoding_format, normalized, etc. are omitted for Ollama batch (Ollama doesn't support these params)
5. ✅ No fallback - Added comment documenting the assumption that /v1/embeddings is universally supported

Rebase Check

Verified: embedder.ts has not changed in upstream master since PR base (16ee3e4). No rebase needed.

Ready for merge! 🎉