Skip to content

LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. Instead of traditional RAG (retrieve-and-answer from scratch every time), the LLM incrementally builds and maintains a persistent wiki from your sources。

TypeScript

fix: support doubao embeddings

nash_su11292ea4e16e
diff --git a/src-tauri/src/commands/cli_resolver.rs b/src-tauri/src/commands/cli_resolver.rs
index 2fad5fbc..725316e1 100644
--- a/src-tauri/src/commands/cli_resolver.rs
+++ b/src-tauri/src/commands/cli_resolver.rs
@@ -22,21 +22,34 @@ static RESOLVED_SHELL_PATH: OnceLock<Option<String>> = OnceLock::new();
/// (cached, so the shell is spawned at most once). Returns `None` when there is
/// nothing to add, in which case the child should inherit PATH unchanged.
#[cfg(not(windows))]
-pub(crate) fn child_path_env() -> Option<String> {
- let shell_path = RESOLVED_SHELL_PATH
- .get_or_init(|| login_shell_path(LOGIN_SHELL_PATH_TIMEOUT))
- .clone()?;
- match std::env::var("PATH") {
- Ok(current) if !current.is_empty() => Some(format!("{shell_path}:{current}")),
- _ => Some(shell_path),
- }
+pub(crate) async fn child_path_env() -> Option<String> {
+ let shell_path = tokio::task::spawn_blocking(|| {
+ RESOLVED_SHELL_PATH
+ .get_or_init(|| login_shell_path(LOGIN_SHELL_PATH_TIMEOUT))
+ .clone()
+ })
+ .await
+ .ok()
+ .flatten()?;
+ Some(merge_child_path_env(
+ &shell_path,
+ std::env::var("PATH").ok().as_deref(),
+ ))
}
#[cfg(windows)]
-pub(crate) fn child_path_env() -> Option<String> {
+pub(crate) async fn child_path_env() -> Option<String> {
None
}
+#[cfg(not(windows))]
+fn merge_child_path_env(shell_path: &str, inherited_path: Option<&str>) -> String {
+ match inherited_path {
+ Some(current) if !current.is_empty() => format!("{shell_path}:{current}"),
+ _ => shell_path.to_string(),
+ }
+}
+
pub(crate) async fn find_cli_command(
command: &str,
windows_candidates: &[&str],
@@ -180,7 +193,7 @@ fn parse_shell_path_output(stdout: &str) -> Option<String> {
#[cfg(all(test, not(windows)))]
mod tests {
- use super::parse_shell_path_output;
+ use super::{merge_child_path_env, parse_shell_path_output};
#[test]
fn parse_shell_path_output_ignores_banners() {
@@ -197,4 +210,24 @@ mod tests {
assert_eq!(parse_shell_path_output("\x1ePATH=\x1e"), None);
assert_eq!(parse_shell_path_output("\x1eOTHER=/usr/bin\x1e"), None);
}
+
+ #[test]
+ fn merge_child_path_env_prepends_shell_path_when_inherited_path_exists() {
+ assert_eq!(
+ merge_child_path_env("/opt/homebrew/bin:/usr/local/bin", Some("/usr/bin:/bin")),
+ "/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin",
+ );
+ }
+
+ #[test]
+ fn merge_child_path_env_uses_shell_path_when_inherited_path_is_empty() {
+ assert_eq!(
+ merge_child_path_env("/opt/homebrew/bin", Some("")),
+ "/opt/homebrew/bin"
+ );
+ assert_eq!(
+ merge_child_path_env("/opt/homebrew/bin", None),
+ "/opt/homebrew/bin"
+ );
+ }
}
diff --git a/src-tauri/src/commands/codex_cli.rs b/src-tauri/src/commands/codex_cli.rs
index de8ef2f4..ba5d5cd0 100644
--- a/src-tauri/src/commands/codex_cli.rs
+++ b/src-tauri/src/commands/codex_cli.rs
@@ -87,7 +87,7 @@ pub async fn codex_cli_detect() -> Result<DetectResult, String> {
// `codex` is a node shim (`#!/usr/bin/env node`); under a GUI launch the
// inherited PATH lacks node, so hand it the login shell PATH or its
// shebang fails with `env: node: No such file or directory`.
- if let Some(path_env) = child_path_env() {
+ if let Some(path_env) = child_path_env().await {
cmd.env("PATH", path_env);
}
let output = tokio::time::timeout(Duration::from_secs(3), cmd.arg("--version").output()).await;
@@ -148,7 +148,7 @@ pub async fn codex_cli_spawn(
suppress_windows_console(&mut cmd);
// See `codex_cli_detect`: the node shim needs the login shell PATH at run
// time so its shebang resolves `node` under a GUI launch.
- if let Some(path_env) = child_path_env() {
+ if let Some(path_env) = child_path_env().await {
cmd.env("PATH", path_env);
}
cmd.args(build_codex_cli_args(&model, isolate_local_config));
diff --git a/src-tauri/src/commands/search.rs b/src-tauri/src/commands/search.rs
index e34f8631..b8e487a7 100644
--- a/src-tauri/src/commands/search.rs
+++ b/src-tauri/src/commands/search.rs
@@ -668,10 +668,11 @@ pub fn extract_image_refs(content: &str) -> Vec<SearchImageRef> {
async fn fetch_embedding(text: &str, cfg: &SearchEmbeddingConfig) -> Result<Vec<f32>, String> {
let is_google = is_google_embedding_config(cfg);
+ let is_doubao_multimodal = is_doubao_multimodal_embedding_config(cfg);
let endpoint = if is_google {
google_embedding_endpoint(cfg)
} else {
- cfg.endpoint.trim().to_string()
+ volcengine_embedding_endpoint(cfg)
};
let mut req = reqwest::Client::builder()
.timeout(std::time::Duration::from_secs(
@@ -703,6 +704,8 @@ async fn fetch_embedding(text: &str, cfg: &SearchEmbeddingConfig) -> Result<Vec<
}
let body = if is_google {
google_embedding_body(&cfg.model, text, cfg.output_dimensionality)
+ } else if is_doubao_multimodal {
+ doubao_multimodal_embedding_body(&cfg.model, text)
} else {
json!({ "model": cfg.model, "input": text })
};
@@ -722,10 +725,22 @@ async fn fetch_embedding(text: &str, cfg: &SearchEmbeddingConfig) -> Result<Vec<
data.to_string().chars().take(200).collect::<String>()
));
}
+ parse_embedding_values(&data, is_google, is_doubao_multimodal)
+}
+
+fn parse_embedding_values(
+ data: &Value,
+ is_google: bool,
+ is_doubao_multimodal: bool,
+) -> Result<Vec<f32>, String> {
let values = if is_google {
data.get("embedding")
.and_then(|v| v.get("values"))
.and_then(Value::as_array)
+ } else if is_doubao_multimodal {
+ data.get("data")
+ .and_then(|v| v.get("embedding"))
+ .and_then(Value::as_array)
} else {
data.get("data")
.and_then(Value::as_array)
@@ -788,6 +803,91 @@ fn is_google_embedding_config(cfg: &SearchEmbeddingConfig) -> bool {
endpoint.contains("generativelanguage.googleapis.com") || endpoint.contains(":embedcontent")
}
+fn is_volcengine_embedding_endpoint(endpoint: &str) -> bool {
+ let host = reqwest::Url::parse(endpoint)
+ .ok()
+ .and_then(|url| url.host_str().map(|host| host.to_lowercase()))
+ .unwrap_or_else(|| {
+ let trimmed = endpoint.trim();
+ trimmed
+ .split_once("://")
+ .map(|(_, rest)| rest)
+ .unwrap_or(trimmed)
+ .split(['/', '?', '#'])
+ .next()
+ .unwrap_or("")
+ .to_lowercase()
+ });
+ host == "volces.com" || host.ends_with(".volces.com") || host.contains("volcengine")
+}
+
+fn is_doubao_multimodal_embedding_config(cfg: &SearchEmbeddingConfig) -> bool {
+ cfg.model
+ .trim()
+ .to_lowercase()
+ .contains("doubao-embedding-vision")
+}
+
+fn volcengine_embedding_endpoint(cfg: &SearchEmbeddingConfig) -> String {
+ let raw = cfg.endpoint.trim();
+ if !is_volcengine_embedding_endpoint(raw) {
+ return raw.to_string();
+ }
+ let suffix = if is_doubao_multimodal_embedding_config(cfg) {
+ "/embeddings/multimodal"
+ } else {
+ "/embeddings"
+ };
+ append_endpoint_path(raw, suffix)
+}
+
+fn append_endpoint_path(endpoint: &str, target_suffix: &str) -> String {
+ let suffix = target_suffix.trim_start_matches('/');
+ match reqwest::Url::parse(endpoint) {
+ Ok(mut url) => {
+ let path = url.path().trim_end_matches('/').to_string();
+ let lower_path = path.to_lowercase();
+ let lower_suffix = format!("/{}", suffix.to_lowercase());
+ if lower_path.ends_with(&lower_suffix) {
+ url.set_path(if path.is_empty() { "/" } else { &path });
+ return url.to_string();
+ }
+ if lower_path.ends_with("/embeddings/multimodal") && lower_suffix == "/embeddings" {
+ let base = path.trim_end_matches("/multimodal");
+ url.set_path(if base.is_empty() { "/" } else { base });
+ return url.to_string();
+ }
+ if lower_path.ends_with("/embeddings") && lower_suffix == "/embeddings/multimodal" {
+ url.set_path(&format!("{path}/multimodal"));
+ return url.to_string();
+ }
+ let next = format!("{path}/{suffix}").replace("//", "/");
+ url.set_path(&next);
+ url.to_string()
+ }
+ Err(_) => {
+ let (base, query) = endpoint.split_once('?').unwrap_or((endpoint, ""));
+ let trimmed = base.trim_end_matches('/');
+ let lower = trimmed.to_lowercase();
+ let lower_suffix = format!("/{}", suffix.to_lowercase());
+ let next = if lower.ends_with(&lower_suffix) {
+ trimmed.to_string()
+ } else if lower.ends_with("/embeddings/multimodal") && lower_suffix == "/embeddings" {
+ trimmed.trim_end_matches("/multimodal").to_string()
+ } else if lower.ends_with("/embeddings") && lower_suffix == "/embeddings/multimodal" {
+ format!("{trimmed}/multimodal")
+ } else {
+ format!("{trimmed}/{suffix}")
+ };
+ if query.is_empty() {
+ next
+ } else {
+ format!("{next}?{query}")
+ }
+ }
+ }
+}
+
fn google_embedding_endpoint(cfg: &SearchEmbeddingConfig) -> String {
let raw = strip_google_api_key_query(cfg.endpoint.trim())
.trim_end_matches('/')
@@ -858,6 +958,14 @@ fn google_embedding_body(model: &str, text: &str, output_dimensionality: Option<
body
}
+fn doubao_multimodal_embedding_body(model: &str, text: &str) -> Value {
+ json!({
+ "model": model,
+ "encoding_format": "float",
+ "input": [{ "type": "text", "text": text }],
+ })
+}
+
pub fn build_snippet(content: &str, query: &str) -> String {
let lower = content.to_lowercase();
let q = query.to_lowercase();
@@ -1007,6 +1115,125 @@ mod tests {
assert_eq!(body["output_dimensionality"], 768);
}
+ #[test]
+ fn volcengine_embedding_endpoint_only_appends_for_volcengine_hosts() {
+ let cfg = SearchEmbeddingConfig {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3".to_string(),
+ api_key: "ARK_KEY".to_string(),
+ model: "doubao-embedding-text-240715".to_string(),
+ output_dimensionality: None,
+ extra_headers: None,
+ };
+ assert_eq!(
+ volcengine_embedding_endpoint(&cfg),
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings"
+ );
+
+ let custom = SearchEmbeddingConfig {
+ endpoint: "https://gateway.example.com/proxy/volcengine?upstream=volces.com"
+ .to_string(),
+ ..cfg.clone()
+ };
+ assert_eq!(
+ volcengine_embedding_endpoint(&custom),
+ "https://gateway.example.com/proxy/volcengine?upstream=volces.com"
+ );
+ }
+
+ #[test]
+ fn doubao_multimodal_endpoint_and_body_use_vision_shape() {
+ let cfg = SearchEmbeddingConfig {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3/embeddings?trace=1".to_string(),
+ api_key: "ARK_KEY".to_string(),
+ model: "doubao-embedding-vision".to_string(),
+ output_dimensionality: None,
+ extra_headers: None,
+ };
+
+ assert_eq!(
+ volcengine_embedding_endpoint(&cfg),
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal?trace=1"
+ );
+
+ let body = doubao_multimodal_embedding_body("doubao-embedding-vision", "hello");
+ assert_eq!(body["model"], "doubao-embedding-vision");
+ assert_eq!(body["encoding_format"], "float");
+ assert_eq!(body["input"][0]["type"], "text");
+ assert_eq!(body["input"][0]["text"], "hello");
+ }
+
+ #[test]
+ fn volcengine_embedding_endpoint_does_not_duplicate_existing_suffixes() {
+ let cfg = SearchEmbeddingConfig {
+ enabled: true,
+ endpoint: "https://ARK.cn-beijing.volces.com/api/v3/embeddings".to_string(),
+ api_key: "ARK_KEY".to_string(),
+ model: "doubao-embedding-text-240715".to_string(),
+ output_dimensionality: None,
+ extra_headers: None,
+ };
+ assert_eq!(
+ volcengine_embedding_endpoint(&cfg),
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings"
+ );
+
+ let vision = SearchEmbeddingConfig {
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal".to_string(),
+ model: "doubao-embedding-vision".to_string(),
+ ..cfg.clone()
+ };
+ assert_eq!(
+ volcengine_embedding_endpoint(&vision),
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal"
+ );
+
+ let text_on_multimodal_endpoint = SearchEmbeddingConfig {
+ model: "doubao-embedding-text-240715".to_string(),
+ ..vision
+ };
+ assert_eq!(
+ volcengine_embedding_endpoint(&text_on_multimodal_endpoint),
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings"
+ );
+ }
+
+ #[test]
+ fn doubao_multimodal_detection_is_model_driven_for_custom_gateways() {
+ let cfg = SearchEmbeddingConfig {
+ enabled: true,
+ endpoint: "https://gateway.example.com/ark/embeddings/multimodal".to_string(),
+ api_key: "KEY".to_string(),
+ model: "doubao-embedding-vision".to_string(),
+ output_dimensionality: None,
+ extra_headers: None,
+ };
+
+ assert!(is_doubao_multimodal_embedding_config(&cfg));
+ assert_eq!(
+ volcengine_embedding_endpoint(&cfg),
+ "https://gateway.example.com/ark/embeddings/multimodal"
+ );
+
+ let body = doubao_multimodal_embedding_body(&cfg.model, "hello");
+ assert_eq!(body["input"][0]["type"], "text");
+ assert_eq!(body["input"][0]["text"], "hello");
+ }
+
+ #[test]
+ fn doubao_multimodal_response_shape_is_distinct_from_openai_shape() {
+ let values =
+ parse_embedding_values(&json!({ "data": { "embedding": [0.1, 0.2] } }), false, true)
+ .expect("vision response should parse");
+ assert_eq!(values, vec![0.1, 0.2]);
+
+ assert!(
+ parse_embedding_values(&json!({ "data": [{ "embedding": [0.1] }] }), false, true,)
+ .is_err()
+ );
+ }
+
#[test]
fn extra_embedding_header_names_are_validated_and_reserved_names_are_skipped() {
assert!(is_safe_extra_header_name("X-Model-Provider-Id"));
diff --git a/src/components/settings/sections/embedding-section.tsx b/src/components/settings/sections/embedding-section.tsx
index 4b05d175..8672e8fd 100644
--- a/src/components/settings/sections/embedding-section.tsx
+++ b/src/components/settings/sections/embedding-section.tsx
@@ -53,6 +53,8 @@ const EMBEDDING_MODEL_SUGGESTIONS = [
"gemini-embedding-001",
"gemini-embedding-2",
"text-embedding-004",
+ "doubao-embedding-vision",
+ "doubao-embedding-text-240715",
"text-embedding-qwen3-embedding-0.6b",
"nomic-embed-text",
"mxbai-embed-large",
diff --git a/src/i18n/en.json b/src/i18n/en.json
index 7ddb5f41..953fd669 100644
--- a/src/i18n/en.json
+++ b/src/i18n/en.json
@@ -302,11 +302,11 @@
},
"embedding": {
"title": "Vector Embeddings",
- "description": "Enable semantic search against your wiki. Works with OpenAI-compatible /v1/embeddings endpoints and Google Gemini native embedContent endpoints. Local models (LM Studio, Ollama embedding) don't need an API key.",
+ "description": "Enable semantic search against your wiki. Works with OpenAI-compatible /v1/embeddings endpoints, Google Gemini native embedContent endpoints, and Volcengine/Doubao embedding endpoints. Local models (LM Studio, Ollama embedding) don't need an API key.",
"enableLabel": "Enable vector search",
"enableHint": "When off, search uses token matching only.",
"endpoint": "Endpoint",
- "endpointHint": "OpenAI-compatible example: http://127.0.0.1:1234/v1/embeddings. Gemini example: https://generativelanguage.googleapis.com/v1beta with model gemini-embedding-001.",
+ "endpointHint": "OpenAI-compatible example: http://127.0.0.1:1234/v1/embeddings. Gemini example: https://generativelanguage.googleapis.com/v1beta with model gemini-embedding-001. Volcengine example: https://ark.cn-beijing.volces.com/api/v3 with model doubao-embedding-vision.",
"apiKey": "API Key (optional)",
"apiKeyPlaceholder": "Leave empty for local models",
"model": "Model",
diff --git a/src/i18n/zh.json b/src/i18n/zh.json
index f84df0cd..273a211e 100644
--- a/src/i18n/zh.json
+++ b/src/i18n/zh.json
@@ -302,11 +302,11 @@
},
"embedding": {
"title": "向量嵌入",
- "description": "启用语义搜索。支持 OpenAI 兼容的 /v1/embeddings endpoint,也支持 Google Gemini 原生 embedContent endpoint。本地模型(如 LM Studio、Ollama embedding)不需要 API Key。",
+ "description": "启用语义搜索。支持 OpenAI 兼容的 /v1/embeddings endpoint、Google Gemini 原生 embedContent endpoint,以及火山方舟 / 豆包 embedding endpoint。本地模型(如 LM Studio、Ollama embedding)不需要 API Key。",
"enableLabel": "启用向量搜索",
"enableHint": "关掉后搜索只走 token 匹配。",
"endpoint": "Endpoint",
- "endpointHint": "OpenAI 兼容示例:http://127.0.0.1:1234/v1/embeddings。Gemini 示例:https://generativelanguage.googleapis.com/v1beta,模型填 gemini-embedding-001。",
+ "endpointHint": "OpenAI 兼容示例:http://127.0.0.1:1234/v1/embeddings。Gemini 示例:https://generativelanguage.googleapis.com/v1beta,模型填 gemini-embedding-001。火山方舟示例:https://ark.cn-beijing.volces.com/api/v3,模型填 doubao-embedding-vision。",
"apiKey": "API Key(可选)",
"apiKeyPlaceholder": "本地模型留空即可",
"model": "Model",
diff --git a/src/lib/embedding.test.ts b/src/lib/embedding.test.ts
index fd8abbae..3b591b9f 100644
--- a/src/lib/embedding.test.ts
+++ b/src/lib/embedding.test.ts
@@ -304,6 +304,187 @@ describe("fetchEmbedding — provider wire formats", () => {
expect(headers["X-Empty"]).toBeUndefined()
})
+ it("does not auto-append /embeddings for generic OpenAI-compatible custom endpoints", async () => {
+ mockHttpFetch.mockResolvedValueOnce(okResponse([0.5, 0.6]))
+
+ const out = await fetchEmbedding("hi", {
+ enabled: true,
+ endpoint: "https://gateway.example.com/v1",
+ apiKey: "sk-test",
+ model: "text-embedding-3-small",
+ })
+
+ expect(out).toEqual([0.5, 0.6])
+ const [url, opts] = mockHttpFetch.mock.calls[0]
+ expect(url).toBe("https://gateway.example.com/v1")
+ expect(JSON.parse(String(opts?.body))).toEqual({
+ model: "text-embedding-3-small",
+ input: "hi",
+ })
+ })
+
+ it("auto-appends /embeddings for Volcengine OpenAI-compatible base endpoints only", async () => {
+ mockHttpFetch.mockResolvedValueOnce(okResponse([0.1, 0.2]))
+
+ const out = await fetchEmbedding("hi", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3",
+ apiKey: "ark-key",
+ model: "doubao-embedding-text-240715",
+ })
+
+ expect(out).toEqual([0.1, 0.2])
+ const [url, opts] = mockHttpFetch.mock.calls[0]
+ expect(url).toBe("https://ark.cn-beijing.volces.com/api/v3/embeddings")
+ expect((opts?.headers as Record<string, string>).Authorization).toBe("Bearer ark-key")
+ expect(JSON.parse(String(opts?.body))).toEqual({
+ model: "doubao-embedding-text-240715",
+ input: "hi",
+ })
+ })
+
+ it("does not infer Volcengine from a custom endpoint path or query string", async () => {
+ mockHttpFetch.mockResolvedValueOnce(okResponse([0.2, 0.3]))
+
+ await fetchEmbedding("hi", {
+ enabled: true,
+ endpoint: "https://gateway.example.com/proxy/volcengine?upstream=volces.com",
+ apiKey: "sk-test",
+ model: "text-embedding-3-small",
+ })
+
+ expect(mockHttpFetch.mock.calls[0][0]).toBe(
+ "https://gateway.example.com/proxy/volcengine?upstream=volces.com",
+ )
+ })
+
+ it("uses Volcengine Doubao multimodal embedding request and response shape", async () => {
+ mockHttpFetch.mockResolvedValueOnce(
+ new Response(JSON.stringify({ data: { embedding: [0.7, 0.8] } }), {
+ status: 200,
+ headers: { "Content-Type": "application/json" },
+ }),
+ )
+
+ const out = await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3",
+ apiKey: "ark-key",
+ model: "doubao-embedding-vision",
+ })
+
+ expect(out).toEqual([0.7, 0.8])
+ const [url, opts] = mockHttpFetch.mock.calls[0]
+ expect(url).toBe("https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal")
+ expect((opts?.headers as Record<string, string>).Authorization).toBe("Bearer ark-key")
+ expect(JSON.parse(String(opts?.body))).toEqual({
+ model: "doubao-embedding-vision",
+ encoding_format: "float",
+ input: [{ type: "text", text: "hello" }],
+ })
+ })
+
+ it("uses Doubao multimodal wire shape for proxied vision models without rewriting custom endpoints", async () => {
+ mockHttpFetch.mockResolvedValueOnce(
+ new Response(JSON.stringify({ data: { embedding: [0.4, 0.5] } }), {
+ status: 200,
+ headers: { "Content-Type": "application/json" },
+ }),
+ )
+
+ const out = await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://gateway.example.com/ark/embeddings/multimodal",
+ apiKey: "proxy-key",
+ model: "doubao-embedding-vision",
+ })
+
+ expect(out).toEqual([0.4, 0.5])
+ const [url, opts] = mockHttpFetch.mock.calls[0]
+ expect(url).toBe("https://gateway.example.com/ark/embeddings/multimodal")
+ expect(JSON.parse(String(opts?.body))).toEqual({
+ model: "doubao-embedding-vision",
+ encoding_format: "float",
+ input: [{ type: "text", text: "hello" }],
+ })
+ })
+
+ it("preserves query parameters when building Volcengine embedding endpoints", async () => {
+ mockHttpFetch.mockResolvedValueOnce(
+ new Response(JSON.stringify({ data: { embedding: [0.9] } }), {
+ status: 200,
+ headers: { "Content-Type": "application/json" },
+ }),
+ )
+
+ await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3/embeddings?trace=1",
+ apiKey: "ark-key",
+ model: "doubao-embedding-vision",
+ })
+
+ expect(mockHttpFetch.mock.calls[0][0]).toBe(
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal?trace=1",
+ )
+ })
+
+ it("does not duplicate existing Volcengine embedding suffixes", async () => {
+ mockHttpFetch
+ .mockResolvedValueOnce(okResponse([0.1]))
+ .mockResolvedValueOnce(
+ new Response(JSON.stringify({ data: { embedding: [0.2] } }), {
+ status: 200,
+ headers: { "Content-Type": "application/json" },
+ }),
+ )
+ .mockResolvedValueOnce(okResponse([0.3]))
+
+ await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3/embeddings",
+ apiKey: "ark-key",
+ model: "doubao-embedding-text-240715",
+ })
+ await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal",
+ apiKey: "ark-key",
+ model: "doubao-embedding-vision",
+ })
+ await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal",
+ apiKey: "ark-key",
+ model: "doubao-embedding-text-240715",
+ })
+
+ expect(mockHttpFetch.mock.calls.map((call) => call[0])).toEqual([
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings",
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings/multimodal",
+ "https://ark.cn-beijing.volces.com/api/v3/embeddings",
+ ])
+ })
+
+ it("reports the Doubao multimodal response shape when the vector is missing", async () => {
+ mockHttpFetch.mockResolvedValueOnce(
+ new Response(JSON.stringify({ data: [{ embedding: [0.1] }] }), {
+ status: 200,
+ headers: { "Content-Type": "application/json" },
+ }),
+ )
+
+ const out = await fetchEmbedding("hello", {
+ enabled: true,
+ endpoint: "https://ark.cn-beijing.volces.com/api/coding/v3",
+ apiKey: "ark-key",
+ model: "doubao-embedding-vision",
+ })
+
+ expect(out).toBeNull()
+ expect(getLastEmbeddingError()).toContain("missing data.embedding")
+ })
+
it("does not let custom headers override the Gemini API key header", async () => {
mockHttpFetch.mockResolvedValueOnce(
new Response(JSON.stringify({ embedding: { values: [0.7, 0.8] } }), {
diff --git a/src/lib/embedding.ts b/src/lib/embedding.ts
index ca99e87f..df8a82f8 100644
--- a/src/lib/embedding.ts
+++ b/src/lib/embedding.ts
@@ -105,7 +105,10 @@ export async function fetchEmbedding(
if (!cfg.endpoint) return null
const isGoogleNative = isGoogleEmbeddingConfig(cfg)
- const endpoint = isGoogleNative ? googleEmbeddingEndpoint(cfg) : cfg.endpoint
+ const isDoubaoMultimodal = isDoubaoMultimodalEmbeddingConfig(cfg)
+ const endpoint = isGoogleNative
+ ? googleEmbeddingEndpoint(cfg)
+ : volcengineEmbeddingEndpoint(cfg)
const headers: Record<string, string> = { "Content-Type": "application/json" }
if (cfg.apiKey) {
if (isGoogleNative) {
@@ -135,7 +138,9 @@ export async function fetchEmbedding(
body: JSON.stringify(
isGoogleNative
? googleEmbeddingBody(cfg.model, current, cfg.outputDimensionality)
- : { model: cfg.model, input: current },
+ : isDoubaoMultimodal
+ ? doubaoMultimodalEmbeddingBody(cfg.model, current)
+ : { model: cfg.model, input: current },
),
})
@@ -143,12 +148,18 @@ export async function fetchEmbedding(
const data = await resp.json()
const embedding = isGoogleNative
? data?.embedding?.values ?? null
- : data?.data?.[0]?.embedding ?? null
+ : isDoubaoMultimodal
+ ? data?.data?.embedding ?? null
+ : data?.data?.[0]?.embedding ?? null
if (isNonEmptyNumberArray(embedding)) {
lastEmbeddingError = null
return embedding
}
- const expectedShape = isGoogleNative ? "embedding.values" : "data[0].embedding"
+ const expectedShape = isGoogleNative
+ ? "embedding.values"
+ : isDoubaoMultimodal
+ ? "data.embedding"
+ : "data[0].embedding"
lastEmbeddingError = `Embedding response missing ${expectedShape} (got ${JSON.stringify(data).slice(0, 200)})`
console.warn(`[Embedding] ${lastEmbeddingError}`)
return null
@@ -215,6 +226,74 @@ function isGoogleEmbeddingConfig(cfg: EmbeddingConfig): boolean {
|| /:embedcontent(\?|$)/i.test(endpoint)
}
+function isVolcengineEmbeddingEndpoint(endpoint: string): boolean {
+ try {
+ const host = new URL(endpoint).hostname.toLowerCase()
+ return host === "volces.com"
+ || host.endsWith(".volces.com")
+ || host.includes("volcengine")
+ } catch {
+ const authority = endpoint
+ .trim()
+ .replace(/^[a-z][a-z0-9+.-]*:\/\//i, "")
+ .split(/[/?#]/, 1)[0]
+ .toLowerCase()
+ return authority === "volces.com"
+ || authority.endsWith(".volces.com")
+ || authority.includes("volcengine")
+ }
+}
+
+function isDoubaoMultimodalEmbeddingConfig(cfg: EmbeddingConfig): boolean {
+ return cfg.model.trim().toLowerCase().includes("doubao-embedding-vision")
+}
+
+function volcengineEmbeddingEndpoint(cfg: EmbeddingConfig): string {
+ const raw = cfg.endpoint.trim()
+ if (!isVolcengineEmbeddingEndpoint(raw)) return raw
+ const targetSuffix = isDoubaoMultimodalEmbeddingConfig(cfg)
+ ? "/embeddings/multimodal"
+ : "/embeddings"
+ return appendEndpointPath(raw, targetSuffix)
+}
+
+function appendEndpointPath(endpoint: string, targetSuffix: string): string {
+ const suffix = targetSuffix.replace(/^\/+/, "")
+ try {
+ const url = new URL(endpoint)
+ const path = url.pathname.replace(/\/+$/, "")
+ const lowerPath = path.toLowerCase()
+ const lowerSuffix = `/${suffix.toLowerCase()}`
+ if (lowerPath.endsWith(lowerSuffix)) {
+ url.pathname = path || "/"
+ return url.toString()
+ }
+ if (lowerPath.endsWith("/embeddings/multimodal") && lowerSuffix === "/embeddings") {
+ url.pathname = path.slice(0, -"/multimodal".length) || "/"
+ return url.toString()
+ }
+ if (lowerPath.endsWith("/embeddings") && lowerSuffix === "/embeddings/multimodal") {
+ url.pathname = `${path}/multimodal`
+ return url.toString()
+ }
+ url.pathname = `${path}/${suffix}`.replace(/\/{2,}/g, "/")
+ return url.toString()
+ } catch {
+ const [base, query = ""] = endpoint.split("?", 2)
+ const trimmed = base.replace(/\/+$/, "")
+ const lower = trimmed.toLowerCase()
+ const lowerSuffix = `/${suffix.toLowerCase()}`
+ const next = lower.endsWith(lowerSuffix)
+ ? trimmed
+ : lower.endsWith("/embeddings/multimodal") && lowerSuffix === "/embeddings"
+ ? trimmed.slice(0, -"/multimodal".length)
+ : lower.endsWith("/embeddings") && lowerSuffix === "/embeddings/multimodal"
+ ? `${trimmed}/multimodal`
+ : `${trimmed}/${suffix}`
+ return query ? `${next}?${query}` : next
+ }
+}
+
function googleEmbeddingEndpoint(cfg: EmbeddingConfig): string {
const raw = stripGoogleApiKeyQuery(cfg.endpoint.trim()).replace(/\/+$/, "")
if (/:batchEmbedContents(\?|$)/i.test(raw)) {
@@ -265,6 +344,14 @@ function googleEmbeddingBody(
return body
}
+function doubaoMultimodalEmbeddingBody(model: string, text: string): Record<string, unknown> {
+ return {
+ model,
+ encoding_format: "float",
+ input: [{ type: "text", text }],
+ }
+}
+
// ── LanceDB v2 operations (via Rust Tauri commands) ──────────────────────
interface ChunkUpsertInput {
diff --git a/src/lib/llm-client.test.ts b/src/lib/llm-client.test.ts
index 8ec0d227..8b7e02c5 100644
--- a/src/lib/llm-client.test.ts
+++ b/src/lib/llm-client.test.ts
@@ -166,4 +166,48 @@ describe("streamChat — mid-stream abort mapping", () => {
expect(onDone).toHaveBeenCalledTimes(1)
expect(onError).not.toHaveBeenCalled()
})
+
+ it("recognises lowercase and single-l cancelled spellings as silent cancels", async () => {
+ for (const message of ["request cancelled", "Request canceled"]) {
+ const { response, getReject, readCalled } = pendingStreamResponse()
+ mockHttpFetch.mockResolvedValueOnce(response)
+
+ const onError = vi.fn()
+ const onDone = vi.fn()
+ const promise = streamChat(
+ cfg,
+ [{ role: "user", content: "hi" }],
+ { onToken: vi.fn(), onDone, onError },
+ undefined,
+ {},
+ )
+
+ await readCalled
+ getReject()(message)
+ await promise
+
+ expect(onDone).toHaveBeenCalledTimes(1)
+ expect(onError).not.toHaveBeenCalled()
+ }
+ })
+
+ it("treats pre-fetch bare-string cancel spellings as silent cancels", async () => {
+ for (const message of ["request cancelled", "Request canceled"]) {
+ mockHttpFetch.mockReset()
+ mockHttpFetch.mockRejectedValueOnce(message)
+
+ const onError = vi.fn()
+ const onDone = vi.fn()
+ await streamChat(
+ cfg,
+ [{ role: "user", content: "hi" }],
+ { onToken: vi.fn(), onDone, onError },
+ undefined,
+ {},
+ )
+
+ expect(onDone).toHaveBeenCalledTimes(1)
+ expect(onError).not.toHaveBeenCalled()
+ }
+ })
})
diff --git a/src/lib/llm-client.ts b/src/lib/llm-client.ts
index 68e8a70f..b49fc306 100644
--- a/src/lib/llm-client.ts
+++ b/src/lib/llm-client.ts
@@ -47,6 +47,11 @@ function parseLines(chunk: Uint8Array, buffer: string): [string[], string] {
return [lines, remaining]
}
+function isRequestCancelledError(err: unknown): boolean {
+ const message = err instanceof Error ? err.message : String(err)
+ return /^request cancel(?:l)?ed$/i.test(message.trim())
+}
+
export async function streamChat(
config: LlmConfig,
messages: import("./llm-providers").ChatMessage[],
@@ -119,7 +124,7 @@ export async function streamChat(
onDone()
return
}
- if (err instanceof Error && err.name === "AbortError") {
+ if ((err instanceof Error && err.name === "AbortError") || isRequestCancelledError(err)) {
// Backstop timeout aborted the request (we tracked this via
// timeoutFired); treat it as a real timeout rather than a cancel.
if (timeoutFired) {
@@ -265,7 +270,7 @@ export async function streamChat(
signal?.aborted ||
timeoutFired ||
(err instanceof Error && err.name === "AbortError") ||
- (err instanceof Error ? err.message : String(err)) === "Request cancelled"
+ isRequestCancelledError(err)
if (isAbort) {
// Mirror the pre-fetch catch: distinguish our long-horizon backstop
// (an actionable timeout) from a user-initiated cancel (silent).