feat(provider): Ollama/local provider + LAT_LLM_BASE_URL/MODEL env overrides

[ikeniborn]

Summary

• LAT_LLM_KEY=ollama — connects to Ollama at http://localhost:11434/v1 using nomic-embed-text (768-dim). No API key needed.
• LAT_LLM_KEY=local — alias for ollama.
• LAT_LLM_BASE_URL — overrides apiBase for any provider (openai, vercel, ollama). Useful for self-hosted OpenAI-compatible endpoints.
• LAT_LLM_MODEL — overrides the embedding model name for any provider.

Motivation

Users running local LLM stacks (Ollama, LM Studio, vLLM, etc.) currently have no way to use lat search without an OpenAI or Vercel API key. The REPLAY_LAT_LLM_KEY::URL workaround hardcodes model: 'replay' which most local servers reject.

Usage

# Ollama (fully local, no API key)
export LAT_LLM_KEY=ollama
ollama pull nomic-embed-text
lat search "architecture overview"

# Custom OpenAI-compatible endpoint
export LAT_LLM_KEY="sk-..."
export LAT_LLM_BASE_URL="https://my-proxy/v1"
export LAT_LLM_MODEL="text-embedding-ada-002"
lat search "authentication flow"

Test plan

• LAT_LLM_KEY=ollama resolves to http://localhost:11434/v1 + nomic-embed-text
• LAT_LLM_BASE_URL overrides apiBase for openai provider
• LAT_LLM_MODEL overrides model for vercel provider
• Existing sk- / vck_ / REPLAY_LAT_LLM_KEY:: paths unchanged
• Error message updated to mention ollama

🤖 Generated with Claude Code