fix: add support for float encoding format in OpenAI Compatible embedder

[roomote]

This PR fixes issue #7180 by adding support for the float encoding format in the OpenAI Compatible embedder.

Problem

Some OpenAI-compatible providers (like certain litellm configurations) only support encoding_format: "float" and not "base64". The current implementation was hardcoded to use "base64", causing errors with these providers.

Solution

• Added an optional encodingFormat parameter to the OpenAICompatibleEmbedder constructor
• Default to "base64" for backward compatibility
• When encodingFormat is set to "float", the embedder:
  • Omits the encoding_format parameter from the OpenAI SDK call (letting it default to float)
  • Uses "float" in direct HTTP requests
  • Handles float array responses without base64 decoding
• Added configuration support in interfaces and config manager
• Added comprehensive tests for both encoding formats

Testing

• Added new test file openai-compatible-encoding.spec.ts with tests for both formats
• All existing tests pass
• Verified backward compatibility with default base64 format

Breaking Changes

None - the change is backward compatible with existing configurations.

Fixes #7180

---

Important

Adds support for 'float' encoding format in OpenAICompatibleEmbedder, with backward compatibility for 'base64'.

• Behavior:
  • Adds encodingFormat parameter to OpenAICompatibleEmbedder constructor, defaulting to "base64".
  • Supports "float" encoding by omitting encoding_format in SDK calls and using it in HTTP requests.
  • Handles float array responses without base64 decoding.
• Configuration:
  • Updates CodeIndexConfigManager to include encodingFormat in openAiCompatibleOptions.
  • Modifies CodeIndexConfig interface to support encodingFormat.
• Testing:
  • Adds openai-compatible-encoding.spec.ts with tests for both encoding formats.
  • Ensures backward compatibility by defaulting to "base64" when encodingFormat is not specified.
• Misc:
  • Updates service-factory.ts to pass encodingFormat to OpenAICompatibleEmbedder.

^{This description was created by for e8df38a. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Reviewing my own code is like debugging in a mirror - everything looks backward but the bugs are still mine.

[roomote]

I notice that in the direct HTTP request path, we always send encoding_format regardless of whether it's "float" or "base64" (line 220). However, in the SDK path (lines 285-287), we only set it for base64. Should we align this behavior for consistency? Some providers might be strict about not accepting the encoding_format parameter when expecting float arrays.

[roomote]

Using (codebaseIndexConfig as any) here bypasses TypeScript's type checking. Consider properly extending the config interface to include codebaseIndexOpenAiCompatibleEncodingFormat for better type safety. This would help catch potential issues at compile time.

[roomote]

The constructor accepts the encodingFormat parameter but doesn't validate it. Consider adding validation to ensure only "base64" or "float" are accepted:

Suggested change

	this.encodingFormat = encodingFormat || "base64"
	// Default to base64 for backward compatibility, but allow float format
	const validFormats = ["base64", "float"] as const
	this.encodingFormat = encodingFormat || "base64"
	if (!validFormats.includes(this.encodingFormat)) {
	throw new Error(`Invalid encoding format: ${encodingFormat}. Must be 'base64' or 'float'`)
	}

[roomote]

Great test coverage! Consider adding a few edge cases:

• What happens when an invalid encoding format like "invalid" is provided?
• How do we handle mixed responses (some base64, some float) from providers?
• Error handling when base64 decoding fails due to corrupted data?

[daniel-lxs]

Duplicate