Benchmarks

We benchmark quality and speed across all methods on ~1,250 queries over 63 repositories in 19 languages.

Main Results

Method	NDCG@10	Index time	Query p50
CodeRankEmbed Hybrid	0.862	57 s	16 ms
semble	0.854	263 ms	1.5 ms
CodeRankEmbed	0.765	57 s	16 ms
ColGREP	0.693	5.8 s	124 ms
BM25	0.673	263 ms	0.02 ms
grepai	0.561	35 s	48 ms
probe	0.387	-	207 ms
ripgrep	0.126	-	12 ms

Semble achieves 99% of the retrieval quality of the 137M-parameter CodeRankEmbed Hybrid, while indexing 218× faster and answering queries 11× faster, entirely on CPU.

The charts below plot latency against NDCG@10. Marker size reflects model parameter count.

Time to first result (index + query) vs NDCG@10

Query latency on a warm index vs NDCG@10

Token Efficiency

Coding agents (Claude Code, OpenCode, etc.) typically find code by running grep on keywords and reading the matched files. We model that workflow and compare it against semble’s chunk retrieval across our full benchmark of 1,251 queries.

Expected tokens per query

For each query: tokens consumed at first relevant hit, or 32k if the method never finds anything. Averaged across all 1,251 queries.

Method	Expected tokens	Savings
ripgrep + read file	45,692	baseline
semble	566	98% fewer

Recall at fixed token budgets

A relevant file is “covered” once any retrieved unit comes from it.

Method	500	1k	2k	4k	8k	16k	32k
semble	0.685	0.849	0.938	0.976	0.991	0.996	0.996
ripgrep + read file	0.001	0.008	0.037	0.088	0.212	0.379	0.583

Methodology

Semble returns the top-50 ranked chunks. ripgrep+read splits the query into keywords (dropping stopwords and short words), runs rg --fixed-strings --ignore-case for each keyword, then reads matched files in full ranked by how many distinct keywords they contain. Both methods search the same set of file types and ignored directories. Tokens are counted with cl100k_base via tiktoken. A relevant file is “covered” once any retrieved unit overlaps its annotated span.

By Language

NDCG@10 per language. Best score per row is bolded.

Language	semble	CRE Hybrid	CRE	ColGREP	grepai	probe	ripgrep
scala	0.909	0.922	0.845	0.765	0.330	0.392	0.180
cpp	0.915	0.913	0.846	0.626	0.731	0.375	0.126
ruby	0.909	0.909	0.769	0.708	0.643	0.382	0.230
elixir	0.894	0.905	0.869	0.808	0.669	0.412	0.134
javascript	0.917	0.903	0.920	0.823	0.675	0.588	0.176
zig	0.913	0.901	0.807	0.474	0.755	0.369	0.000
csharp	0.885	0.889	0.743	0.614	0.277	0.392	0.117
go	0.895	0.884	0.676	0.785	0.722	0.410	0.133
python	0.867	0.880	0.794	0.777	0.634	0.488	0.202
php	0.858	0.874	0.758	0.663	0.402	0.340	0.123
swift	0.860	0.873	0.721	0.710	0.429	0.280	0.160
bash	0.825	0.852	0.892	0.706	0.723	0.226	0.000
lua	0.823	0.847	0.803	0.798	0.699	0.336	0.000
java	0.849	0.841	0.706	0.641	0.386	0.536	0.198
kotlin	0.821	0.830	0.670	0.637	0.478	0.335	0.166
rust	0.856	0.827	0.627	0.662	0.519	0.242	0.162
c	0.741	0.806	0.706	0.676	0.555	0.384	0.000
haskell	0.765	0.771	0.776	0.683	0.483	0.313	0.000
typescript	0.706	0.708	0.545	0.430	0.394	0.354	0.128
overall	0.854	0.862	0.765	0.693	0.561	0.387	0.126

Ablations

raw returns retrieval scores directly; + ranking feeds them through semble’s hybrid reranker.

Retrieval	Raw	+ ranking
BM25	0.675	0.834
potion-code-16M	0.650	0.821
BM25 + potion-code-16M	-	0.854

By query category:

Mode	Architecture	Semantic	Symbol
BM25 raw	0.628	0.676	0.719
potion-code-16M raw	0.626	0.666	0.629
semble BM25 (+ ranking)	0.770	0.819	0.957
semble potion-code-16M (+ ranking)	0.757	0.808	0.943
semble hybrid	0.802	0.846	0.958

Dataset

~1,250 queries over 63 repositories in 19 languages, grouped into three categories:

Category	Queries	What it tests
semantic	711	Code that implements a specific behavior or concept
architecture	343	Design decisions, module boundaries, structural patterns
symbol	204	Named entity lookup (function, class, type, variable)

Languages covered: bash, C, C++, C#, Elixir, Go, Haskell, Java, JavaScript, Kotlin, Lua, PHP, Python, Ruby, Rust, Scala, Swift, TypeScript, Zig.

Methods

• ripgrep: fast regex search, included as a raw keyword-match baseline.
• probe: BM25 keyword ranking backed by tree-sitter parse trees. No persistent index; scans on the fly.
• ColGREP: late-interaction code retrieval with the LateOn-Code-edge model.
• grepai: semantic search using nomic-embed-text (137M params) via a local Ollama daemon.
• CodeRankEmbed: 137M-param transformer embedding model. CRE Hybrid fuses its dense scores with BM25.
• semble: potion-code-16M static embeddings + BM25 + the semble reranking stack.