Embedding models

The aligner compares sentences by their embedding vectors, so the embedding model decides how many translations are found. Three local GGUF models are configured out of the box. All three run through llama.cpp, with Metal on Apple Silicon.

profile	model	prompt prefix	weights licence
jina (default)	jina-embeddings-v5-text-nano, text-matching, Q8_0	Query:	CC BY-NC
gemma	EmbeddingGemma 300M, Q8_0	task: sentence similarity \| query:	see the model card
granite	granite-embedding-311m-multilingual-r2, Q8_0	none	Apache 2.0

Select a model per call with --model, or set [embed] model in paraltext.toml. Point each profile at the file on your disk with [embed.models.<name>] path (see Configuration).

Which one to use

The models were compared on whole books from a private collection of translated books: Stephen King's On Writing (15 editions), Marshall McLuhan's Understanding Media (13), Alberto Manguel's A History of Reading (10) and Lakoff and Johnson's Metaphors We Live By (6). All runs used English as pivot, threshold 1.05 and centred vectors.

	jina	gemma	granite
Strengths	Most units that span all editions; never collapses on a language; fastest.	2-8% more pairs on major Western languages and Chinese.	Most even across languages; best on Czech, Hungarian and Lithuanian; permissive licence.
Weaknesses	Non-commercial licence on the weights.	Lithuanian yield of 32-42%; weak on Hungarian.	About 7% fewer pairs overall; slowest.

Where two models link the same pivot sentence, they pick the same translation 93-100% of the time. The choice of model mostly changes how many sentences are found, not which translation is picked.

Rules of thumb:

• Use jina when you want units across many editions, or when the set includes smaller languages.
• Use gemma when all editions are in major Western languages or Chinese and recall matters.
• Use granite when the weights must carry a permissive licence, or for Baltic and Finno-Ugric languages.

King, 15 editions

	jina	gemma	granite
links over 14 target languages	41,625	41,274	38,901
best on	da, ro, sv	de, es, it, pl, pt, ru, tr, zh	cs, hu, lt
units in all 15 editions	176	98	127
units in 10 or more editions	2,362	2,348	2,064
embedding 127k texts	559 s	608 s	701 s

McLuhan, 13 editions

	jina	gemma	granite
links over 12 target languages	54,349	55,128	51,994
Lithuanian yield	53%	42%	56%
units in all 13 editions	422	325	346
units in 10 or more editions	2,971	3,120	2,652

The pattern repeats: gemma edges the totals through the large languages, jina holds every language and wins the all-edition units by 20-30%.

Why vectors are centred

Granite's vector space is compressed: random sentence pairs across languages score a cosine of about 0.80 and true translations about 0.97. The best neighbour is then never 5% better than the average neighbour, so the ratio margin never reaches 1.05. On the Lakoff book, Granite found 145 English-Spanish links without centring.

Subtracting the joint mean vector of both editions removes the shared offset. With [embed] center = true, the default, Granite found 2,578 English-Spanish links on the same book, and jina and gemma gained 4-12% as well.

Other backends

The uubed engine also runs profiles with other backends. The example configuration declares local SentenceTransformers and MLX profiles (nomic, minilm, voyage-nano) and API profiles (jina-api, gemma-api, granite-api, voyage-lite). The API backends read credentials from environment variables. The comparisons above cover only the three GGUF models.

Precision and cache size

Vectors are stored in the cache at [embed] precision. int8 is the default. In a test run on one EPUB book, float32 and int8 vectors produced 63 units each; 62 were identical and one alignment changed.