Embedding-aware Topological Indexing for Your Architecture

, the default retrieval engine in Eulix

ETIYA is one of the two retrieval engines in eulix-cli, and the default. It runs the same exact-match, keyword and call-graph strategies as EHNAS, and adds a search over embeddings so a question can reach code that shares none of its words.

Full name
Embedding-aware Topological Indexing for Your Architecture
Part of
eulix-cli (Go)
Status
Beta, and the default engine
Needs
embeddings.bin and vectors.bin from eulix analyze, and the local embedding helper at query time

What ETIYA does

Every question first goes through the strategies both engines share: exact symbol matches, partial matches, keywords and the call graph.

ETIYA also embeds the question through the local embedding helper and compares it with the stored chunk vectors by cosine similarity, using an IVF index on large projects. That finds code for questions worded nothing like the code. The semantic step is skipped for callers and callees questions, which the call graph answers exactly.

How it selects context

Candidates are picked with query-aware MMR (Maximal Marginal Relevance): each pick weighs relevance to the question against overlap with what is already chosen, so the context window isn’t filled with near-duplicates.

If embeddings.bin is missing, ETIYA falls back to keyword-only retrieval and says so in the debug log, because a silent fallback would look exactly like a quality regression.

ETIYA or EHNAS?

ETIYA is the default, and the engine existing projects were built and tuned with. Pick EHNAS for huge repositories, for machines with no network or embedder, or when you need scores you can recompute exactly. Both read the same chunk store, so switching is one line in eulix.toml.

Turn it on

In your project’s eulix.toml:

[project]
engine = "etiya"

ETIYA documentation

More of Eulix

All Eulix docs