fix: decode base64-encoded embeddings in recorder

[iskhakovt]

Problem

When recording fixtures, the recorder fails to save OpenAI embedding responses that use base64 encoding. The OpenAI API returns base64-encoded embeddings when encoding_format: "base64" is set (the default in many SDKs, including Python's openai package).

buildFixtureResponse checks Array.isArray(first.embedding) but the embedding is a base64 string, not an array. The fixture is saved with a proxy_error:

{
  "response": {
    "error": { "message": "Could not detect response format from upstream", "type": "proxy_error" },
    "status": 200
  }
}

Fix

Extract encoding_format from the raw request body and pass it to buildFixtureResponse. When encoding_format === "base64" and the embedding is a string, decode it via Buffer.from(embedding, "base64") → Float32Array → Array.from(floats).

This is forward-safe — only decodes when encoding_format explicitly says "base64", so future encoding formats (e.g. "int8") won't be misinterpreted.

Changes

• recorder.ts: Extract encoding_format from raw request, pass to buildFixtureResponse, decode base64 when detected

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/CopilotKit/llmock/@copilotkit/llmock@64

commit: 5602fb5

[jpr5]

Thanks for the fix! The base64 decoding logic is correct — good use of Float32Array with byteOffset for buffer pool alignment, and gating on encoding_format === "base64" is the right approach.

I've added 3 tests covering the 2×2 matrix of (encoding_format present/absent) × (base64 string / array embedding):

1. base64 string + encoding_format: "base64" → decodes to [0.5, 1, -0.25]
2. base64 string + no encoding_format → falls through to proxy_error (the original bug path)
3. array embedding + encoding_format: "base64" → Array.isArray wins, array passthrough

All red-green verified. The tests are on our branch `fix/base64-embedding-recording` — feel free to cherry-pick commit `0617f69` or I can merge them into this PR.

One minor note: if `buf.byteLength % 4 !== 0` (corrupted base64), the `Float32Array` constructor will throw an uncaught `RangeError`. Not a blocker since OpenAI won't send corrupted data, but a `try/catch` around the decode with fallthrough to the existing error path would be more defensive.