Dhee's context summarisation and embedding pipeline sends data to an external cloud LLM. This means every file read, edit, and conversation turn processed by dhee is potentially summarised via a third-party server. The privacy.py regex filter catches credentials, but proprietary code, internal architecture, and business logic still egress. For developers in regulated industries or working on sensitive codebases, this is a hard blocker — and it undermines dhee's core value proposition of being a private developer brain.
Proposal
Add an optional local provider that runs a lightweight on-device model (e.g. smollm2-135M or Qwen2.5-0.5B via llama-cpp-python for summarisation, and all-MiniLM-L6-v2 via sentence-transformers for embeddings) so that no data leaves the machine. These models are small enough (~90–400 MB) to download on first use to ~/.dhee/models/ and run comfortably on CPU or Apple Silicon. Ideally, local-first would become the default on dhee init, with cloud providers as an explicit opt-in — matching the trust expectations of a tool that handles your most sensitive developer context.
Dhee's context summarisation and embedding pipeline sends data to an external cloud LLM. This means every file read, edit, and conversation turn processed by dhee is potentially summarised via a third-party server. The
privacy.pyregex filter catches credentials, but proprietary code, internal architecture, and business logic still egress. For developers in regulated industries or working on sensitive codebases, this is a hard blocker — and it undermines dhee's core value proposition of being a private developer brain.Proposal
Add an optional local provider that runs a lightweight on-device model (e.g.
smollm2-135MorQwen2.5-0.5Bviallama-cpp-pythonfor summarisation, andall-MiniLM-L6-v2viasentence-transformersfor embeddings) so that no data leaves the machine. These models are small enough (~90–400 MB) to download on first use to~/.dhee/models/and run comfortably on CPU or Apple Silicon. Ideally, local-first would become the default ondhee init, with cloud providers as an explicit opt-in — matching the trust expectations of a tool that handles your most sensitive developer context.