Local AI assistant in Zed: RAG + web browsing with pluggable models

An OSS effort aims to add local-first AI to the Zed IDE: in-IDE web browsing, workspace embeddings into a vector DB (e.g., Qdrant), semantic code search, and cross-file refactors using Ollama or OpenAI-compatible APIs. The author is seeking architectural guidance on AST-aware chunking, incremental re-indexing for polyglot repos, efficient streaming to the editor, sandboxed web fetching with prompt-injection defenses, and model orchestration (router vs task-specific). A Zed fork already exposes settings for vector DB URL, embedder provider/model, API keys, and an index toggle.