ll-core 0.1.2

Hybrid semantic search engine: BM25 + vector + PageRank + cross-encoder reranking over SQLite, with on-device ONNX inference
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67

use anyhow::Result;
use serde::{Deserialize, Serialize};

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ModelConfig {
    pub model_id: String,
    pub dim: usize,
    pub max_tokens: usize,
    pub query_prefix: Option<String>,
    pub passage_prefix: Option<String>,
    pub needs_token_type_ids: bool,
    pub needs_external_pooling: bool,
    pub normalize_embeddings: bool,
    pub output_tensor_name: Option<String>,
}

pub trait EmbeddingProvider: Send + Sync {
    fn embed_batch(&self, texts: &[String]) -> Result<Vec<Vec<f32>>>;
    fn config(&self) -> &ModelConfig;

    fn embed_query(&self, text: &str) -> Result<Vec<f32>> {
        let prefixed = match &self.config().query_prefix {
            Some(prefix) => format!("{}{}", prefix, text),
            None => text.to_string(),
        };
        let mut results = self.embed_batch(&[prefixed])?;
        let vec = results.remove(0);
        let expected = self.config().dim;
        anyhow::ensure!(
            vec.len() == expected,
            "embedding dim mismatch: model {} produced {} dims, expected {}",
            self.config().model_id, vec.len(), expected
        );
        Ok(vec)
    }

    fn embed_documents(&self, texts: &[String]) -> Result<Vec<Vec<f32>>> {
        if texts.is_empty() {
            return Ok(Vec::new());
        }
        let results = match &self.config().passage_prefix {
            Some(prefix) => {
                let prefixed: Vec<String> =
                    texts.iter().map(|t| format!("{}{}", prefix, t)).collect();
                self.embed_batch(&prefixed)?
            }
            None => self.embed_batch(texts)?,
        };
        let expected = self.config().dim;
        if let Some(vec) = results.first() {
            anyhow::ensure!(
                vec.len() == expected,
                "embedding dim mismatch: model {} produced {} dims, expected {}",
                self.config().model_id, vec.len(), expected
            );
        }
        Ok(results)
    }

    fn dim(&self) -> usize {
        self.config().dim
    }

    fn model_id(&self) -> &str {
        &self.config().model_id
    }
}