Feature request: support OpenAI-compatible LLM APIs (e.g. DeepSeek) and custom local embedding providers

[sirenexcelsior]

Summary

It would be very helpful if XMem added:

1. First-class support for OpenAI-compatible LLM APIs such as DeepSeek
2. A configurable custom local embedding provider for self-hosted embedding services

Why this would help

XMem is especially attractive for local-first and self-hosted multi-agent setups. In those environments, users often already have:

• lower-cost or self-hosted OpenAI-compatible chat APIs
• local embedding services such as BGE-M3
• existing private AI infrastructure they want XMem to plug into

Right now, DeepSeek support is indirect and local embedding support is limited to a fixed set of provider shapes. That makes adoption harder than it needs to be.

Proposal

1. Add a generic OpenAI-compatible LLM provider

Example config:

OPENAI_COMPATIBLE_API_KEY=...
OPENAI_COMPATIBLE_BASE_URL=https://api.deepseek.com/v1
OPENAI_COMPATIBLE_MODEL=deepseek-chat

This would support:

• DeepSeek official API
• other OpenAI-compatible vendors
• self-hosted OpenAI-compatible gateways

2. Add a generic custom embedding HTTP provider

Example config:

EMBEDDING_PROVIDER=custom
CUSTOM_EMBEDDING_BASE_URL=http://127.0.0.1:8001/embed
CUSTOM_EMBEDDING_MODEL=bge-m3
PINECONE_DIMENSION=1024

Suggested behavior:

• send text to a user-configured embedding endpoint
• accept a simple embedding response shape
• validate returned vector dimension against PINECONE_DIMENSION
• keep the rest of the pipeline unchanged

Example use case

A personal multi-agent setup where:

• XMem runs locally
• xmem-mcp is used by Codex, Claude Code, and Hermes
• DeepSeek is used for LLM calls
• local BGE-M3 is used for embeddings

This feels like a very natural XMem deployment model, but today it requires indirect routing or code patches.

Suggested acceptance criteria

• XMem can call a configurable OpenAI-compatible LLM endpoint using API key, base URL, and model from config
• XMem can call a configurable local embedding HTTP endpoint
• embedding dimension validation still works as expected
• existing providers continue to work without breaking changes
• docs include at least one example for DeepSeek-compatible LLM setup and one example for custom local embeddings

Thanks for considering this — it would make XMem much easier to adopt in personal and self-hosted agent workflows.