Struct JinaLateChunkingClient 

pub struct JinaLateChunkingClient { /* private fields */ }

Jina AI embeddings client with native late chunking support

Calls the Jina embeddings API with late_chunking=true. The API encodes the concatenated inputs as a single sequence and returns per-input embeddings where each embedding reflects the full-document context.

For fully local operation (no API key), configure Ollama with rjmalagon/gte-qwen2-7b-instruct — it provides excellent 32k-context embeddings without native late chunking but with a far larger context window.

Examples

use graphrag_core::text::late_chunking::JinaLateChunkingClient;

let client = JinaLateChunkingClient::new("jina_xxxx".to_string());
let embeddings = client.embed_with_late_chunking(chunks).await?;
assert_eq!(embeddings.len(), chunks.len());

Implementations

impl JinaLateChunkingClient

pub fn new(api_key: impl Into<String>) -> Self

Create a new client with a Jina API key

pub fn with_model(self, model: impl Into<String>) -> Self

Override the embedding model

pub async fn embed_with_late_chunking( &self, chunks: &[TextChunk], ) -> Result<Vec<Vec<f32>>>

Embed chunks using Jina’s late chunking API

Sends all chunk contents in a single request with late_chunking: true. The Jina API encodes them as one sequence so each chunk’s embedding incorporates the full document context.

Returns one embedding vector per chunk, in the same order as the input.

Trait Implementations

impl Clone for JinaLateChunkingClient

fn clone(&self) -> JinaLateChunkingClient

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for JinaLateChunkingClient

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.