Exact · Hierarchical · Node · Structure
, retrieval without embeddings in Eulix
EHNAS is the second retrieval engine in eulix-cli. It meets the same need as vector search, reaching code that shares no words with your question, without running a model at query time. Every score in its core path can be recomputed from source metadata.
- Full name
- Exact · Hierarchical · Node · Structure
- Part of
- eulix-cli (Go)
- Status
- Beta
- Needs
- No embeddings, no GPU and no outbound calls at query time
Why it exists
Vector search is the usual bridge between a question and code that uses different words, but why a chunk ranks high is then a fact about a model’s internals. EHNAS was built to meet the same need deterministically, so “why was this retrieved?” always has an answer you can check.
Three layers on a deterministic core
The core is deterministic: exact matching, a path gate, subsystem boosting and scope demotion. At the scale of millions of lines, EHNAS adds one layer for each problem the core can’t solve:
- Twin search. Chunks with the same structural signature in another language are scored together, so logic ported between languages is found even with no shared words.
- Monte Carlo walk. 256 walkers start from the strongest hits and move over symbol, call-graph and word links. Common symbols end walks sooner, and the cost per query stays about the same whatever the repository size.
- Novelty packing. As the context window fills, blocks that repeat symbols already picked count for less, so the window isn’t spent on near-duplicate boilerplate.
How it selects context
EHNAS selects with structural MMR: redundancy is measured from the code itself (shared symbols, nearby lines) rather than from the question. Each layer reports its contribution separately in the debug trace, so the deterministic part of any score can always be recovered.
When to pick EHNAS
Huge repositories, machines with no network or embedder, and cases where you need reproducible scoring. With EHNAS, chat and mcp start without launching the embedding helper.
Turn it on
In your project’s eulix.toml:
[project]
engine = "ehnas"EHNAS documentation
- Retrieval engines: ETIYA and EHNASHow the two engines differ and how to choose.
- EHNAS augmentation layersTwin search, the walk, novelty packing and structural MMR in detail.
- Search pipelineThe deterministic core both engines share.
- Retrieval overviewThe whole retrieval pipeline.
- ConfigThe engine setting and per-engine overrides.
- Retrieval and query issuesKnown problems and workarounds.