dsh-context-milvus

Encoding agents inside large codebases hit a familiar wall: string-based grep cannot find semantically related code with different names, multi-round tool calls flood the context with irrelevant text, token costs spike, and model reasoning degrades. dsh-context-milvus plugs a Milvus vector database directly into the DSH Agent process. Code is chunked along function, class, and method boundaries using tree-sitter AST parsing, embedded, and stored in Milvus so that retrieval shifts from needle-in-a-haystack grep to precise recall of relevant snippets, delivering clean context in a single tool call.

On core capabilities, the plugin offers hybrid retrieval that fuses BM25 keyword search with vector semantic search via RRF, incremental indexing tracked with Merkle SHA-256 hashes, cross-file import-resolved call chain tracing (find_callers and trace_call_chain), a three-layer gitignore-style ignore system, and an ADR decision memory layer that records the design rationale behind code changes with semantic search, constraint injection, and consistency checks. Spec document fusion extends unified retrieval to brainstorming design docs as well.

It is built for teams running DeepSeek Harness coding agents on mid-to-large repositories: thousands of files, complex cross-module dependencies, a need to reduce tool-call rounds and token overhead, and a desire for the agent to perform impact analysis and understand historical design decisions before making changes. Supports self-hosted Milvus or Zilliz Cloud, keeping data under your own control.