OpenRouter adapter and Voyage AI semantic search added in one commit

Chris-o-O dropped OpenRouter multi-model support and a pgvector semantic search pipeline in a single 925-line commit. Each is gated by its own env var and follows the existing provider abstractions, so they can be adopted independently.

searchinfrastructure

The OpenRouter adapter (backend/src/lib/llm/openrouter.ts) uses the OpenAI SDK pointed at https://openrouter.ai/api/v1 to route to GPT-4o, GPT-4 Turbo, Mistral Large, Llama 3.3 70B, and Gemma 3 27B. It wires into the existing streamChat interface, handles tool-call accumulation across stream chunks, and stores the user's key in a new openrouter_api_key column on user_profiles. The account settings page handles it identically to how Claude and Gemini keys are already handled.

The semantic search pipeline uses Voyage AI's voyage-law-2 model (1024-dimensional embeddings, tuned for legal text). Documents are chunked at 800 characters with 150-character overlap at upload time, fire-and-forget, with [Page N] markers preserved through the chunking. Chunks go into a document_chunks table; a match_document_chunks Supabase RPC does cosine similarity lookup. The search is exposed via a /search/semantic endpoint, a semantic_search chat tool, and a widget on the project page.

Before relying on this in production: indexing has no retry and no backfill, so existing documents won't appear in semantic results until re-uploaded. The pgvector migration requires the extension to be manually enabled in the Supabase dashboard first. The RPC uses security definer and filters by a p_user_id parameter rather than auth.uid(), which is correct for service-role backend usage but means table-level RLS isn't load-bearing here.

So what Worth a look if you want to add OpenRouter as a provider or replace keyword search with vector similarity. Both pieces fit the existing abstractions cleanly. Plan for a backfill script before shipping semantic search to users who have already uploaded documents.

Spotted something wrong? Or know the PR text has fresher detail than the writeup above?

Commits in this thread

1 commit from Chris-o-O/mike-explor, oldest first. Source extracted verbatim from the harvested git log.

SHA	Subject	Author	Date
`3279a584`	Add OpenRouter multi-model support and Voyage AI semantic search	Claude	2026-05-07	↗ GitHub
commit body OpenRouter: extends the LLM provider abstraction to support GPT-4o, GPT-4 Turbo, Mistral Large, Llama 3.3 70B, and Gemma 3 27B via a single OpenRouter API key. Follows the same provider pattern as Claude and Gemini - new openrouter.ts adapter, model routing by prefix, per-user API key stored in user_profiles, and full UI integration (model toggle, availability checks, account settings key field). Semantic search: adds Voyage AI embeddings (voyage-law-2, optimised for legal documents) to index document chunks at upload time (fire-and-forget, controlled by VOYAGE_API_KEY env var). Chunks are stored in a new document_chunks pgvector table. A /search/semantic endpoint and a semantic_search LLM tool enable concept-level retrieval. A new SemanticSearchWidget in ProjectPage exposes document search by meaning. https://claude.ai/code/session_01NJxa5NasJ9qhaMkzEyqr2u

SHA

Subject

Author

Date

3279a584

Add OpenRouter multi-model support and Voyage AI semantic search

Claude

2026-05-07

↗ GitHub

commit body

OpenRouter: extends the LLM provider abstraction to support GPT-4o,
GPT-4 Turbo, Mistral Large, Llama 3.3 70B, and Gemma 3 27B via a
single OpenRouter API key. Follows the same provider pattern as Claude
and Gemini - new openrouter.ts adapter, model routing by prefix, per-user
API key stored in user_profiles, and full UI integration (model toggle,
availability checks, account settings key field).

Semantic search: adds Voyage AI embeddings (voyage-law-2, optimised for
legal documents) to index document chunks at upload time (fire-and-forget,
controlled by VOYAGE_API_KEY env var). Chunks are stored in a new
document_chunks pgvector table. A /search/semantic endpoint and a
semantic_search LLM tool enable concept-level retrieval. A new
SemanticSearchWidget in ProjectPage exposes document search by meaning.

https://claude.ai/code/session_01NJxa5NasJ9qhaMkzEyqr2u

Capture this thread into my fork

Download a single Markdown prompt that tells Claude how to port every commit above into your working tree — adapting paths and structure to match your repo. Run it via claude -p < capture-thread-210.md from inside the repo you want the changes in.