Gemini embeddings return HTTP response objects instead of numeric vectors

Environment

• Zotero GPT: 3.1.8
• Zotero: 9.0.1
• Embedding provider: Gemini / google
• Embedding model: gemini-embedding-2
• API: https://generativelanguage.googleapis.com/v1beta

Problem

When using custom Gemini embeddings in AskPDF / related text search, similarity search fails with an error similar to:

cosine-similarity()::invalid input argument. First argument must be an array.
Value: [object XMLHttpRequest]

Cause

The Gemini/google embedding provider appears to return the raw Zotero.HTTP.request(...) results from Promise.all(...) instead of extracting numeric embedding vectors.

In the bundled 3.1.8 script, the google provider sends requests to Gemini :embedContent, but then returns the HTTP response objects directly. As a result, cosine similarity receives XMLHttpRequest/HTTP response objects instead of number[] vectors.

Expected behavior

The Gemini embedding provider should return number[][], one numeric vector per input text, compatible with OpenAI-compatible embedding providers and cosine similarity.

Suggested fix

Parse the JSON response and extract a numeric vector before returning. Gemini/OpenAI-compatible response shapes may include:

• response.embedding.values
• response.embedding
• response.embeddings[0].values
• response.embeddings[0].embedding.values
• response.embeddings[0].embedding
• response.data[0].embedding

Then return something like:

return responses.map(extractVector);

instead of returning the raw HTTP responses.

For gemini-embedding-2, it may also be safer not to send the old taskType field unless the API explicitly supports it for that model.

Notes

OpenAI-compatible embeddings such as Qwen/Qwen3-Embedding-8B work through the data[].embedding format. The issue seems specific to the Gemini/google embedding branch.