fix: pass encoding_format="float" in LiteLLM SDK embedding calls

[franchb]

Summary

• Explicitly pass encoding_format="float" in all litellm.embedding() / litellm.aembedding() calls in LiteLLMSDKEmbeddings
• DeepInfra (and other OpenAI-compatible providers) reject requests when encoding_format is null/None
• LiteLLM defaults encoding_format to None to prevent the OpenAI SDK from using "base64", but this None value causes 400 errors on providers that validate the field

Problem

When using DeepInfra embedding models (e.g., Qwen/Qwen3-Embedding-8B, BAAI/bge-small-en-v1.5) via the litellm-sdk provider, requests fail with:

BadRequestError: Error code: 400 - 'encoding_format' only support with [float, base64]

This happens because:

1. LiteLLM explicitly sets encoding_format=None when not specified (source) to override OpenAI SDK's default of "base64"
2. DeepInfra's API validates the field and rejects null — it only accepts "float" or "base64"

This is a known issue across the ecosystem: microsoft/graphrag#2194, BerriAI/litellm#12110, RooCodeInc/Roo-Code#7199.

Fix

Pass encoding_format="float" explicitly. This is always safe because LiteLLMSDKEmbeddings.encode() returns list[list[float]] — float is the only compatible format.

Test plan

• Unit tests updated to verify encoding_format="float" is passed
• Run cd hindsight-api && uv run pytest tests/test_litellm_sdk_embeddings.py -v

🤖 Generated with Claude Code

[nicoloboschi]

LGTM