Refresh VoyageAI embedding model catalog - #2

VERDICT:CHANGES_NEEDED

Reviewed the diff, commit hygiene, and fit with the existing docs.

Commit hygiene — clean. Author/committer is fzowl <zoltan@voyageai.com>, no Co-Authored-By, no "Generated with" trailers, and no .claude/ files anywhere in the tree or the diff.

Table refresh — looks good. The voyage-4 family, voyage-context-4, voyage-multimodal-3.5, and the gap-filling rows are formatted consistently with the existing tables, and the dimensions line up with prior conventions (e.g. voyage-3.5-lite = 512 matches the row already on the integration page). Docs-only, so no breaking-change or test concerns.

What needs to change before merge:

1. The contextualized-models {% hint %} note asserts Epsilla-internal behavior that needs maintainer verification. It states specific implementation details — "Epsilla sends the batch as a flat list with auto-chunking enabled and a large chunk size ... on the query/retrieval path auto-chunking is disabled." These are claims about how Epsilla itself processes voyage-context-* inputs, not about the Voyage API. Coming from a Voyage-side contributor, they should be confirmed against the actual Epsilla implementation before being published as fact. Relatedly, as written the note says cross-input contextualization is not applied and each string resolves to a single chunk — i.e. the models behave like ordinary embedding models with no contextual benefit. That directly contradicts the stated purpose of voyage-context-*; if that's genuinely the intended behavior, please make the tradeoff explicit; if not, the note is misleading. (The two files also word the note slightly differently — worth aligning.)

2. Table sync is incomplete / leaves pre-existing duplicates. vector-database/embeddings.md still lists voyage-code-2, voyage-law-2, voyage-finance-2, and voyage-multilingual-2 twice. These predate this PR, but since the stated goal is to bring the two tables in sync, it'd be a good opportunity to de-dupe them here.

Point 1 is the blocker — happy to approve once the behavior note is verified (or dropped) by an Epsilla maintainer.

Addressed the review in 94e7e90.

1. voyage-context- behavior note.* Dropped the internal-processing assertions (flat-list / auto-chunking / chunk-size / query-path mechanics) that couldn't be published as fact without maintainer verification. The note now describes only user-observable behavior and makes the tradeoff explicit: these are contextualized models, but in Epsilla each input string is embedded independently as its own document (one deterministic vector per input) and contextualization across the other texts in a batch is not applied, since a batch may contain unrelated documents. Both platform/integrations/voyage-ai.md and vector-database/embeddings.md now carry identical wording.

2. Table de-dup. Removed the pre-existing duplicate voyage-code-2, voyage-law-2, voyage-finance-2, and voyage-multilingual-2 rows in vector-database/embeddings.md so the two model tables stay in sync.

Model catalog and dimensions were re-checked against the current Voyage AI embeddings and contextualized-embeddings docs; no further table changes needed.

VERDICT:APPROVED

Reviewed the diff against origin/main (base ebcd592), and the PR is in good shape. GitHub reports it as MERGEABLE / CLEAN.

Correctness

• The voyage-4 family, voyage-context-4, and voyage-multimodal-3.5 are added to both tables, and the two tables are now back in sync — a real drift that existed because New VoyageAI models in docs epsilla-cloud/docs#4 only touched the integration page.
• The de-dupe is correct: origin/main's vector-database/embeddings.md had genuine duplicate rows (voyage-code-2, voyage-law-2, voyage-finance-2, voyage-multilingual-2 each appeared twice). After this PR each model appears exactly once (verified — no remaining dupes).
• Dimensions all use the 1024 default, matching the existing row convention.

Fit with conventions

• The {% hint style="info" %} block matches GitBook hint usage already present across ~13 docs pages, so it renders natively.
• The note wording is identical on both pages and, per the second commit, was trimmed to user-observable behavior only (one deterministic vector per input, no cross-batch contextualization) — good call dropping the unverified internal-processing details.

Breaking-change risk: None — docs-only, clean merge.

Tests: N/A for a docs repo; correctly called out in the PR description.

Commit hygiene: Clean. Two commits, both authored by zoltan@voyageai.com. No Co-Authored-By, no "Generated with" trailers, no .claude/ files in the tree.

Non-blocking nit: voyage-3.5-lite is listed at 512. This just mirrors what the integration page already had (from epsilla-cloud#4), so it's consistent and not introduced here — but Voyage's voyage-3.5-lite default output dimension is 1024 (512 is the older voyage-3-lite). Worth a quick double-check against docs.voyageai.com; can be a follow-up either way.

Approving.