FastEmbed Embeddings (Local)

FastEmbed Embeddings (Local) turns text into a numeric fingerprint of its meaning using a model that runs inside the NappAI platform itself. With that fingerprint, NappAI can find the documents that say something similar to what you are looking for, even when they use different words, for example in a knowledge base.

You connect the component to a vector database, which uses it both to store your documents and to search them. It needs no credential or account with any provider, and the text is not sent to any outside service.

When to use it

• Give an agent the knowledge in your manuals, contracts or FAQs by storing them in a vector database.
• Build an internal search that finds documents by their meaning, not only by exact words.
• You want to start without signing up for any AI service.
• Your documents are confidential and you prefer not to send them to an outside provider.
• Your documents are in Spanish and English: the default model is designed for those two languages.

Configuration

Field	What it does
Model	The model that calculates the fingerprints. Default: jinaai/jina-embeddings-v2-base-es. See the table of models below.

Model	What it suits
jinaai/jina-embeddings-v2-base-es	Spanish and English, including long texts. It is the default and the best choice for documents in these two languages.
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	About 50 languages. It is the lightest and fastest model, but the one with the fewest hits when searching.
intfloat/multilingual-e5-large	About 100 languages. Search quality very close to the default model, at the cost of being about three times heavier and slower. Its texts are embedded as they are, without the query: / passage: prefixes its authors suggest.
Snowflake/snowflake-arctic-embed-l-v2.0	About 75 languages, including long texts. It puts the chunk that answers first most often, in exchange for being about three times heavier and slower than the default.

Every model in the list allows commercial use.

Advanced options

Field	What it does
Document Embed Type	default or passage. passage embeds documents with the model’s own document mode, if it has one; with the models in the list both give the same result. No text prefix is added. Default: default.
Max Sequence Length	The maximum length of each text; anything beyond it is cut. Default: 512.
Batch Size	How many texts are processed together. Default: 256.
ONNX Runtime Threads	How many calculation threads are used at the same time. With 0 (the default) it is decided automatically.
Model Cache Directory	Leave it empty.

Output

• Embeddings: the model that calculates the fingerprints, ready to use. Connect it to the Embedding input of a vector database: Astra DB, Cassandra, Chroma DB, Couchbase, Elasticsearch, FAISS, Milvus, MongoDB Vector Store, PGVector, Pinecone, Qdrant, Redis, Supabase Vector Store, Upstash, Vectara, Weaviate Cloud (v4), Neo4j. Also the Embedding input of Neo4j GraphRAG. Other components with an input of this type accept it too, such as Semantic Text Splitter or Text Embedder.

Use the same model to store and to search

Fingerprints from two different models cannot be compared with each other. Use the same component and the same model to store the documents and to search them. If you change the model, store the documents in the database again.

Example

An estate agency, Inmobiliaria Portal Sur, wants its assistant to answer using the sheets of its properties in Spanish and English.

1. In a loading flow, split the documents into chunks with Recursive Character Text Splitter and pass them to Chroma DB with Operation set to Add.
2. Add FastEmbed Embeddings (Local) and connect its Embeddings output to the Embedding input of Chroma DB.
3. In the assistant flow, use the same Chroma DB collection with the same FastEmbed Embeddings (Local) component to find the chunks that answer each question.

Measured figures

Measured on 17 and 18 October 2026 on a Spanish test knowledge base (9 documents of a fictional repair service, 40 questions). It is a single synthetic set: it shows how the models behave, not a promise for your data.

Model	Answer as the first result	Answer in the top 5 results	Size
jinaai/jina-embeddings-v2-base-es	60 %	83 %	0.64 GB
intfloat/multilingual-e5-large	60 %	87 % (one more question)	2.24 GB
Snowflake/snowflake-arctic-embed-l-v2.0	70 %	83 %	2.27 GB
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	23 %	63 %	0.22 GB

The default model finds the answer in the top 5 almost as often as the large ones while being about three times smaller and faster. That is why it is the recommended one for documents in Spanish. If you pass few results to the agent and want the right one first, try Snowflake/snowflake-arctic-embed-l-v2.0.

Related components

• Chroma DB
• PGVector
• Text Embedder
• Embedding Similarity
• Local Rerank: reorders results with another model that also runs inside NappAI.

To choose a model and a vector database, see Vector databases in the RAG guide.