feat(semantic-cache): add Google AI (Gemini) embedding provider

[Vswaroop04]

Summary

Adds Google AI embedding support to both packages, following the same provider pattern already used for Cohere and Voyage.

Added providers

• packages/semantic-cache/src/embed/google.ts
• TypeScript provider using native fetch
• packages/semantic-cache-py/betterdb_semantic_cache/embed/google.py
• Python provider using httpx

Both providers target the embedContent REST API and support:

• Default model: text-embedding-004(768 dimensions)
• Configurable taskType
• defaults to RETRIEVAL_QUERY
• use RETRIEVAL_DOCUMENT when storing documents
• Optional title support for improved retrieval quality with document embeddings
• Optional outputDimensionality truncation (text-embedding-004 and newer)
• API key from GOOGLE_API_KEY or explicit config
• No dependency on the official Google SDK

---

Usage

TypeScript

import { createGoogleEmbed } from '@betterdb/semantic-cache/embed/google';

const embed = createGoogleEmbed({
  model: 'text-embedding-004'
});

const cache = new SemanticCache({
  client,
  embedFn: embed
});

Python

from betterdb_semantic_cache.embed.google import create_google_embed

embed = create_google_embed(
    model="text-embedding-004"
)

cache = SemanticCache(
    SemanticCacheOptions(
        client=client,
        embed_fn=embed
    )
)

---

Test

• pnpm test passes in packages/semantic-cache
• Manually verified against a live GOOGLE_API_KEY
• createGoogleEmbed()("hello world") returns a valid 768-element float embedding array

[Vswaroop04]

@KIvanow @jamby77 would appreciate any feedback or thoughts on the approach here

[KIvanow]

Hey @Vswaroop04 thank you for your contribution! A few issues though:

• Unsigned commit (259e5f1) - same as your other three open PRs.
• API key passed as a URL query parameter - inconsistent with every other provider and a security regression.
  Every existing provider (Cohere, Voyage, OpenAI, etc.) passes the key in a header:
  // cohere.ts / voyage.ts
  headers: { Authorization: Bearer ${apiKey} }
• The Google implementation puts the key in the URL:
  // google.ts - don't do this
  ${baseUrl}/models/${model}:embedContent?key=${encodeURIComponent(apiKey)}
• And Python:
  params={"key": key}
• Keys in URLs end up in server access logs, proxy logs, and HTTP Referer headers. Google's API supports x-goog-api-key as a request header - please use that instead, consistent with how every other provider in this package works.
• httpx.AsyncClient is never closed (Python). The lazy-initialized client in _client[0] holds a connection pool with no cleanup path. In test environments this produces ResourceWarnings; in production it leaks file descriptors over time. Either expose a close() helper or document that callers are responsible for cleanup, but the current code silently leaks.
• title behaviour is inconsistent between TS and Python. TypeScript skips title when falsy (if (opts?.title)), so an empty string is silently dropped. Python sends it when title is not None, so "" would be included in the request body. Pick one semantics and apply it to both - !== undefined / is not None is the more predictable choice.

[Vswaroop04]

Hey @Vswaroop04 thank you for your contribution! A few issues though:

* Unsigned commit ([259e5f1](https://github.com/BetterDB-inc/monitor/commit/259e5f1eaf77c943d5238fdaea42c762cb752087)) - same as your other three open PRs.

* API key passed as a URL query parameter - inconsistent with every other provider and a security regression.
  Every existing provider (Cohere, Voyage, OpenAI, etc.) passes the key in a header:
  // cohere.ts / voyage.ts
  headers: { Authorization: `Bearer ${apiKey}` }

* The Google implementation puts the key in the URL:
  // google.ts - don't do this
  `${baseUrl}/models/${model}:embedContent?key=${encodeURIComponent(apiKey)}`

* And Python:
  params={"key": key}

* Keys in URLs end up in server access logs, proxy logs, and HTTP Referer headers. Google's API supports x-goog-api-key as a request header - please use that instead, consistent with how every other provider in this package works.

* httpx.AsyncClient is never closed (Python). The lazy-initialized client in _client[0] holds a connection pool with no cleanup path. In test environments this produces ResourceWarnings; in production it leaks file descriptors over time. Either expose a close() helper or document that callers are responsible for cleanup, but the current code silently leaks.

* title behaviour is inconsistent between TS and Python. TypeScript skips title when falsy (if (opts?.title)), so an empty string is silently dropped. Python sends it when title is not None, so "" would be included in the request body. Pick one semantics and apply it to both - !== undefined / is not None is the more predictable choice.

Thanks for the thorough review! All issues addressed Should have caught these before opening nice finds!

[KIvanow]

Hey @Vswaroop04 - all four issues from the first review are fixed, thanks! One small thing before this merges:

The close() helper is a good addition but it's invisible to callers — nothing in the module docstring or create_google_embed docstring mentions it exists. Please add a note, e.g.:

  When finished, release the connection pool:
      await embed.close()

One line in the docstring is enough. After that this is good to go.

[Vswaroop04]

Hey @Vswaroop04 - all four issues from the first review are fixed, thanks! One small thing before this merges:

The close() helper is a good addition but it's invisible to callers — nothing in the module docstring or create_google_embed docstring mentions it exists. Please add a note, e.g.:

  When finished, release the connection pool:
      await embed.close()

One line in the docstring is enough. After that this is good to go.

Done