Support local embeddings for self-hosted installs (model downloaded on demand) #202

Why

Without an embedding API key, FrontLit falls back to hash vectors, which barely capture meaning, so self-hosters without a provider key effectively get keyword-only search. A small local embedding model closes that gap and gives a strong self-hosting story:

• Content is never sent to a third party just to be indexed (data residency, regulated customers)
• Works on installs with no internet after setup
• Reindexing a large site has no per-document bill

Reference: Software vs model local RAG uses FastEmbed with a small model on the same machine.

Proposal

• Add a local embedding provider (e.g. FastEmbed / ONNX with a small model such as BGE-small) alongside the existing API providers.
• Do not bundle the model in the Docker image. The model is downloaded when the user configures indexing in their workspace, then cached on a persistent volume.
• When no API key is set, offer local embeddings as the default instead of the hash fallback.
• Hosted FrontLit can keep its current provider.

Details to handle

• Download progress and errors shown in workspace settings; indexing waits until the model is ready (keyword search keeps working meanwhile)
• Configurable model cache path, and an option to pre-seed the model for air-gapped installs
• Store the embedding model id and dimension with the index; changing models triggers a full reindex (vectors from different models can't be mixed)
• Batching and CPU limits so indexing a large site doesn't starve the web server

Acceptance criteria

• Docker image size unchanged (no model files baked in)
• Choosing local embeddings in workspace settings downloads the model once and reuses it across restarts
• Self-hosted install with no API key gets semantic results from the local model, not hash vectors
• Switching models reindexes cleanly
• Retrieval eval suite for help desk and chatbot answers #198 eval shows recall vs the hash fallback

Related

• Hybrid RAG over docs, pages, blog, changelog and API reference (#194) #196 hybrid search
• Retrieval eval suite for help desk and chatbot answers #198 retrieval eval suite