feat: Add adaptors for various backends (ollama, tgi, api-inference)

crates/llm-ls/src/adaptors.rs

@@ -0,0 +1,163 @@
+use super::{
+    internal_error, APIError, APIResponse, CompletionParams, Generation, Ide, NAME, VERSION,
+};
+use reqwest::header::{HeaderMap, HeaderValue, AUTHORIZATION, USER_AGENT};
+use serde::{Deserialize, Serialize};
+use serde_json::Value;
+use tower_lsp::jsonrpc;
+
+fn build_tgi_body(prompt: String, params: CompletionParams) -> Value {
+    serde_json::json!({
+        "inputs": prompt,
+        "parameters": params.request_params,
+    })
+}
+
+fn build_tgi_headers(api_token: Option<&String>, ide: Ide) -> Result<HeaderMap, jsonrpc::Error> {
+    let mut headers = HeaderMap::new();
+    let user_agent = format!("{NAME}/{VERSION}; rust/unknown; ide/{ide:?}");
+    headers.insert(
+        USER_AGENT,
+        HeaderValue::from_str(&user_agent).map_err(internal_error)?,
+    );
+
+    if let Some(api_token) = api_token {
+        headers.insert(
+            AUTHORIZATION,
+            HeaderValue::from_str(&format!("Bearer {api_token}")).map_err(internal_error)?,
+        );
+    }
+
+    Ok(headers)
+}
+
+fn parse_tgi_text(text: reqwest::Result<String>) -> Result<Vec<Generation>, jsonrpc::Error> {
+    let generations =
+        match serde_json::from_str(&text.unwrap_or_default()).map_err(internal_error)? {
+            APIResponse::Generation(gen) => vec![gen],
+            APIResponse::Generations(_) => {
+                return Err(internal_error(
+                    "TGI parser unexpectedly encountered api-inference",
+                ))
+            }
+            APIResponse::Error(err) => return Err(internal_error(err)),
+        };
+    Ok(generations)
+}
+
+fn build_api_body(prompt: String, params: CompletionParams) -> Value {
+    build_tgi_body(prompt, params)
+}
+
+fn build_api_headers(api_token: Option<&String>, ide: Ide) -> Result<HeaderMap, jsonrpc::Error> {
+    build_tgi_headers(api_token, ide)
+}
+
+fn parse_api_text(text: reqwest::Result<String>) -> Result<Vec<Generation>, jsonrpc::Error> {
+    let generations =
+        match serde_json::from_str(&text.unwrap_or_default()).map_err(internal_error)? {
+            APIResponse::Generation(gen) => vec![gen],
+            APIResponse::Generations(gens) => gens,
+            APIResponse::Error(err) => return Err(internal_error(err)),
+        };
+    Ok(generations)
+}
+
+fn build_ollama_body(prompt: String, params: CompletionParams) -> Value {
+    let request_body = params.request_body.unwrap_or_default();
+    let body = serde_json::json!({
+        "prompt": prompt,
+        "model": request_body.get("model"),
+    });
+    body
+}
+fn build_ollama_headers() -> Result<HeaderMap, jsonrpc::Error> {
+    let headers = HeaderMap::new();
+    Ok(headers)
+}
+
+#[derive(Debug, Serialize, Deserialize)]
+struct OllamaGeneration {
+    response: String,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(untagged)]
+enum OllamaAPIResponse {
+    Generation(OllamaGeneration),
+    Error(APIError),
+}
+
+fn parse_ollama_text(
+    text: Result<String, reqwest::Error>,
+) -> Result<Vec<Generation>, jsonrpc::Error> {
+    match text {
+        Ok(text) => {
+            let mut gen: Vec<Generation> = Vec::new();
+            for row in text.split('\n') {
+                if row.is_empty() {
+                    continue;
+                }
+                let chunk = match serde_json::from_str(row) {
+                    Ok(OllamaAPIResponse::Generation(ollama_gen)) => ollama_gen.response,
+                    Ok(OllamaAPIResponse::Error(err)) => return Err(internal_error(err)),
+                    Err(err) => return Err(internal_error(err)),
+                };
+                gen.push(Generation {
+                    generated_text: chunk,
+                })
+            }
+            Ok(gen)
+        }
+        Err(err) => Err(internal_error(err)),
+    }
+}
+
+const TGI: &str = "tgi";
+const HUGGING_FACE: &str = "huggingface";
+const OLLAMA: &str = "ollama";
+const DEFAULT_ADAPTOR: &str = HUGGING_FACE;
+
+fn unknown_adaptor_error(adaptor: String) -> jsonrpc::Error {
+    internal_error(format!("Unknown adaptor {}", adaptor))
+}
+
+pub fn adapt_body(prompt: String, params: CompletionParams) -> Result<Value, jsonrpc::Error> {
+    let adaptor = params
+        .adaptor
+        .clone()
+        .unwrap_or(DEFAULT_ADAPTOR.to_string());
+    match adaptor.as_str() {
+        TGI => Ok(build_tgi_body(prompt, params)),
+        HUGGING_FACE => Ok(build_api_body(prompt, params)),
+        OLLAMA => Ok(build_ollama_body(prompt, params)),
+        _ => Err(unknown_adaptor_error(adaptor)),
+    }
+}
+
+pub fn adapt_headers(
+    adaptor: Option<String>,
+    api_token: Option<&String>,
+    ide: Ide,
+) -> Result<HeaderMap, jsonrpc::Error> {
+    let adaptor = adaptor.clone().unwrap_or(DEFAULT_ADAPTOR.to_string());
+    match adaptor.as_str() {
+        TGI => build_tgi_headers(api_token, ide),
+        HUGGING_FACE => build_api_headers(api_token, ide),
+        OLLAMA => build_ollama_headers(),
+        _ => Err(internal_error(adaptor)),
+    }
+}
+
+pub fn adapt_text(
+    adaptor: Option<String>,
+    text: Result<String, reqwest::Error>,
+) -> jsonrpc::Result<Vec<Generation>> {
+    let adaptor = adaptor.clone().unwrap_or(DEFAULT_ADAPTOR.to_string());
+    match adaptor.as_str() {
+        TGI => parse_tgi_text(text),
+        HUGGING_FACE => parse_api_text(text),
+        OLLAMA => parse_ollama_text(text),
+        _ => Err(unknown_adaptor_error(adaptor)),
+    }
+}

crates/llm-ls/src/main.rs

@@ -1,3 +1,4 @@
+use adaptors::{adapt_body, adapt_headers, adapt_text};
 use document::Document;
 use reqwest::header::{HeaderMap, HeaderValue, AUTHORIZATION, USER_AGENT};
 use ropey::Rope;
@@ -18,12 +19,13 @@ use tracing_appender::rolling;
 use tracing_subscriber::EnvFilter;
 use uuid::Uuid;
 
+mod adaptors;
 mod document;
 mod language_id;
 
 const MAX_WARNING_REPEAT: Duration = Duration::from_secs(3_600);
-const NAME: &str = "llm-ls";
-const VERSION: &str = env!("CARGO_PKG_VERSION");
+pub const NAME: &str = "llm-ls";
+pub const VERSION: &str = env!("CARGO_PKG_VERSION");
 
 fn get_position_idx(rope: &Rope, row: usize, col: usize) -> Result<usize> {
     Ok(rope.try_line_to_char(row).map_err(internal_error)?
@@ -120,7 +122,7 @@ fn should_complete(document: &Document, position: Position) -> Result<Completion
     }
 }
 
-#[derive(Debug, Deserialize, Serialize)]
+#[derive(Debug, Deserialize, Serialize, Clone)]
 #[serde(untagged)]
 enum TokenizerConfig {
     Local { path: PathBuf },
@@ -178,12 +180,12 @@ struct APIRequest {
 }
 
 #[derive(Debug, Serialize, Deserialize)]
-struct Generation {
+pub struct Generation {
     generated_text: String,
 }
 
 #[derive(Debug, Deserialize)]
-struct APIError {
+pub struct APIError {
     error: String,
 }
 
@@ -195,7 +197,7 @@ impl Display for APIError {
 
 #[derive(Debug, Deserialize)]
 #[serde(untagged)]
-enum APIResponse {
+pub enum APIResponse {
     Generation(Generation),
     Generations(Vec<Generation>),
     Error(APIError),
@@ -219,7 +221,7 @@ struct Completion {
 
 #[derive(Clone, Copy, Debug, Default, Deserialize, Serialize)]
 #[serde(rename_all = "lowercase")]
-enum Ide {
+pub enum Ide {
     Neovim,
     VSCode,
     JetBrains,
@@ -261,7 +263,7 @@ struct RejectedCompletion {
 
 #[derive(Debug, Deserialize, Serialize)]
 #[serde(rename_all = "camelCase")]
-struct CompletionParams {
+pub struct CompletionParams {
     #[serde(flatten)]
     text_document_position: TextDocumentPositionParams,
     request_params: RequestParams,
@@ -271,10 +273,12 @@ struct CompletionParams {
     fim: FimParams,
     api_token: Option<String>,
     model: String,
+    adaptor: Option<String>,
     tokens_to_clear: Vec<String>,
     tokenizer_config: Option<TokenizerConfig>,
     context_window: usize,
     tls_skip_verify_insecure: bool,
+    request_body: Option<serde_json::Map<String, serde_json::Value>>,
 }
 
 #[derive(Debug, Deserialize, Serialize)]
@@ -283,7 +287,7 @@ struct CompletionResult {
     completions: Vec<Completion>,
 }
 
-fn internal_error<E: Display>(err: E) -> Error {
+pub fn internal_error<E: Display>(err: E) -> Error {
     let err_msg = err.to_string();
     error!(err_msg);
     Error {
@@ -398,37 +402,34 @@ fn build_prompt(
 
 async fn request_completion(
     http_client: &reqwest::Client,
-    ide: Ide,
-    model: &str,
-    request_params: RequestParams,
-    api_token: Option<&String>,
     prompt: String,
+    params: CompletionParams,
 ) -> Result<Vec<Generation>> {
     let t = Instant::now();
+    let model = params.model.clone();
+    let adaptor = params.adaptor.clone();
+    let api_token = params.api_token.clone();
+    let ide = params.ide;
+
+    let json = adapt_body(prompt, params);
+    let headers = adapt_headers(adaptor.clone(), api_token.as_ref(), ide)?;
     let res = http_client
-        .post(build_url(model))
-        .json(&APIRequest {
-            inputs: prompt,
-            parameters: request_params.into(),
-        })
-        .headers(build_headers(api_token, ide)?)
+        .post(build_url(&model))
+        .json(&json)
+        .headers(headers)
         .send()
         .await
         .map_err(internal_error)?;
 
-    let generations = match res.json().await.map_err(internal_error)? {
-        APIResponse::Generation(gen) => vec![gen],
-        APIResponse::Generations(gens) => gens,
-        APIResponse::Error(err) => return Err(internal_error(err)),
-    };
+    let generations = adapt_text(adaptor, res.text().await);
     let time = t.elapsed().as_millis();
     info!(
         model,
         compute_generations_ms = time,
         generations = serde_json::to_string(&generations).map_err(internal_error)?,
         "{model} computed generations in {time} ms"
     );
-    Ok(generations)
+    generations
 }
 
 fn parse_generations(
@@ -624,10 +625,11 @@ impl Backend {
                 return Ok(CompletionResult { request_id, completions: vec![]});
             }
 
+            let tokenizer_config = params.tokenizer_config.clone();
             let tokenizer = get_tokenizer(
                 &params.model,
                 &mut *self.tokenizer_map.write().await,
-                params.tokenizer_config,
+                tokenizer_config,
                 &self.http_client,
                 &self.cache_dir,
                 params.api_token.as_ref(),
@@ -648,17 +650,15 @@ impl Backend {
             } else {
                 &self.http_client
             };
+            let tokens_to_clear = params.tokens_to_clear.clone();
             let result = request_completion(
                 http_client,
-                params.ide,
-                &params.model,
-                params.request_params,
-                params.api_token.as_ref(),
                 prompt,
+                params,
             )
             .await?;
 
-            let completions = parse_generations(result, &params.tokens_to_clear, completion_type);
+            let completions = parse_generations(result, &tokens_to_clear, completion_type);
             Ok(CompletionResult { request_id, completions })
         }.instrument(span).await
     }
@@ -849,3 +849,4 @@ async fn main() {
 
     Server::new(stdin, stdout, socket).serve(service).await;
 }
+