Voice, Local AI & Agent Tooling
neus Search repositories and make constrained decisions
Hybrid Go and Markdown search with cited answers, bounded classification and local inference.
go install ubunatic.com/neus/cmd/neus@latestWhy neus
Why neus
Agents need to find relevant code when their query uses different words. Programs consuming model decisions also need a predictable output shape and an explicit set of allowed labels.
What it does
What it does
Hybrid repository search
Incremental indexing extracts Go and Markdown, combines keyword retrieval with vector similarity, and groups ranked results by file.
Bounded classification
The classify CLI validates input, returns JSON, and short-circuits exact and prefix matches before calling an inference endpoint.
Local inference
HTTP services or an on-demand llama-server process provide inference while the Go binary avoids a bundled Python ML runtime.
The workflow
The workflow
- Go / Markdown sources
- SQLite text + vectors
- File-grouped retrieval
- Cited answers / bounded labels
Recent milestones
Recent milestones
- 2026-09-27 — Batched embeddings reached 29.5 chunks/s on the recorded corpus; issue 030 closed. 979fba9
- 2026-10-02 — Local fallback and GBNF-constrained classifier work completed in issues 036–037. 5bd5bc2
- 2026-10-03 — The classify CLI exposed deterministic matching and endpoint-backed classification. b4b582f
Historical milestones from the review through October 4, 2026.
How the work was verified
How the work was verified
A small reasoning model exhausted its response budget without reliable structured output. The recorded experiment moved to an instruct model and grammar constraints. The benchmark separates structural validity from semantic accuracy, and studies preserve retrieval approaches that regressed recall.
The October 4 review records 215 commits in its six-week window, 14 Go test files, and 124 test functions. These historical counts show the pace and breadth of search, evaluation, CLI, and model-lifecycle work.
Current scope and limits
Current scope and limits
Vector search and classification require configured inference services or local model weights. Grammar-valid output does not establish semantic accuracy. The October 4 review records open integration bugs around cloud fallback model naming and selecting a model that is not loaded (issues 042–043).
Build the CLI
Build the CLI
Build and installation targets are maintained in the repository Makefile.
Search and classification additionally need configured inference services or local model weights.