bge_m3_embedding_server/handler/
both.rs

1// Copyright (c) 2026 J. Patrick Fulton
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15//! `POST /v1/embeddings:both` handler — dense + sparse embeddings in one pass.
16
17use std::sync::Arc;
18use std::time::Instant;
19
20use axum::{Json, extract::State, http::HeaderMap};
21
22use super::common::{check_ready, collect_x_headers, validate_input};
23use crate::error::AppError;
24use crate::models::{DualEmbeddingData, DualRequest, DualResponse, SparseValues, Usage};
25use crate::state::AppState;
26
27/// Handles `POST /v1/embeddings:both` — returns dense and sparse embeddings in one pass.
28///
29/// # Errors
30///
31/// - [`AppError::ServiceUnavailable`] if the model is not ready or no workers are live.
32/// - [`AppError::InvalidRequest`] if the batch is empty, exceeds `max_batch`, or any
33///   text exceeds the per-string character limit.
34/// - [`AppError::Internal`] if the embedding pool returns an inference error.
35///
36/// # Panics
37///
38/// Panics if the request semaphore has been closed — should not occur in normal operation.
39#[allow(clippy::cast_possible_truncation)]
40#[tracing::instrument(
41    skip(state, req, headers),
42    fields(
43        batch_size,
44        prompt_tokens,
45        chunks,
46        max_chunk_seq,
47        tokenize_ms,
48        inference_ms,
49        queue_wait_ms,
50        total_ms,
51    )
52)]
53pub async fn both_embeddings(
54    State(state): State<Arc<AppState>>,
55    headers: HeaderMap,
56    Json(req): Json<DualRequest>,
57) -> Result<Json<DualResponse>, AppError> {
58    check_ready(&state)?;
59    let x_headers = collect_x_headers(&headers);
60    let texts = req.input.0;
61    drop(req.model);
62    validate_input(&texts, state.max_batch)?;
63    let batch_size = texts.len();
64    tracing::Span::current().record("batch_size", batch_size);
65
66    let prompt_tokens: usize = texts.iter().map(|t| t.chars().count() / 4 + 1).sum();
67    tracing::Span::current().record("prompt_tokens", prompt_tokens);
68
69    let t0 = Instant::now();
70
71    let _permit = Arc::clone(&state.request_permits)
72        .acquire_owned()
73        .await
74        .expect("request semaphore is never closed");
75
76    let queue_wait_ms = u64::try_from(t0.elapsed().as_millis()).unwrap_or(u64::MAX);
77
78    let (pairs, embed_stats) = state.pool.both(texts).await?;
79
80    let total_ms = u64::try_from(t0.elapsed().as_millis()).unwrap_or(u64::MAX);
81    tracing::Span::current()
82        .record("chunks", embed_stats.chunks)
83        .record("max_chunk_seq", embed_stats.max_chunk_seq)
84        .record("tokenize_ms", embed_stats.tokenize_ms)
85        .record("inference_ms", embed_stats.inference_ms)
86        .record("queue_wait_ms", queue_wait_ms)
87        .record("total_ms", total_ms);
88    let x_headers_val =
89        (!x_headers.is_empty()).then(|| serde_json::to_string(&x_headers).unwrap_or_default());
90    tracing::info!(
91        route = "both",
92        batch_size,
93        prompt_tokens,
94        chunks = embed_stats.chunks,
95        max_chunk_seq = embed_stats.max_chunk_seq,
96        total_token_positions = embed_stats.total_token_positions,
97        seq_len_min = embed_stats.seq_len_min,
98        seq_len_max = embed_stats.seq_len_max,
99        seq_len_mean = embed_stats.seq_len_mean,
100        seq_len_p95 = embed_stats.seq_len_p95,
101        tokenize_ms = embed_stats.tokenize_ms,
102        inference_ms = embed_stats.inference_ms,
103        queue_wait_ms,
104        total_ms,
105        x_headers = x_headers_val,
106        "embedding request complete"
107    );
108
109    let data = pairs
110        .into_iter()
111        .enumerate()
112        .map(|(index, pair)| DualEmbeddingData {
113            index,
114            embedding: pair.dense,
115            sparse_values: SparseValues {
116                indices: pair.sparse.indices.iter().map(|i| *i as u32).collect(),
117                values: pair.sparse.values,
118            },
119        })
120        .collect();
121
122    Ok(Json(DualResponse {
123        object: "list",
124        model: "bge-m3",
125        data,
126        usage: Usage {
127            prompt_tokens,
128            total_tokens: prompt_tokens,
129        },
130    }))
131}
bge_m3_embedding_server/handler/both.rs

bge_m3_embedding_server/handler/
both.rs
