Files
windmill/backend/windmill-api/src/ai.rs
T
centdix 8b77fc11d8 chore(aichat): send keepalives to avoid timeout on long requests (#7813)
* send keepalive

* cancel pending tools on error

* refactor(ai): extract keepalive interval to constant

Co-authored-by: centdix <centdix@users.noreply.github.com>

---------

Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: centdix <centdix@users.noreply.github.com>
2026-02-05 17:37:02 +00:00

918 lines
33 KiB
Rust

#[cfg(feature = "bedrock")]
use crate::bedrock;
use crate::db::{ApiAuthed, DB};
#[cfg(feature = "bedrock")]
use axum::routing::get;
#[cfg(feature = "bedrock")]
use axum::Json;
use axum::{body::Bytes, extract::Path, response::IntoResponse, routing::post, Extension, Router};
use futures::StreamExt;
use http::{HeaderMap, Method};
use quick_cache::sync::Cache;
use reqwest::{Client, RequestBuilder};
use serde::{Deserialize, Serialize};
use serde_json::{json, value::RawValue};
use std::collections::HashMap;
use std::time::Duration;
use windmill_audit::{audit_oss::audit_log, ActionKind};
use windmill_common::ai_providers::{
empty_string_as_none, AIProvider, ProviderConfig, ProviderModel,
};
use windmill_common::error::{to_anyhow, Error, Result};
use windmill_common::utils::configure_client;
use windmill_common::variables::get_variable_or_self;
// AI timeout configuration constants
const AI_TIMEOUT_MIN_SECS: u64 = 1;
const AI_TIMEOUT_MAX_SECS: u64 = 86400; // 24 hours
const AI_TIMEOUT_DEFAULT_SECS: u64 = 3600; // 1 hour
const HTTP_POOL_MAX_IDLE_PER_HOST: usize = 10;
const HTTP_POOL_IDLE_TIMEOUT_SECS: u64 = 90;
const KEEPALIVE_INTERVAL_SECS: u64 = 15;
lazy_static::lazy_static! {
/// AI request timeout in seconds.
///
/// This timeout applies to the TOTAL duration of AI HTTP requests,
/// including streaming responses. Default is 3600 seconds (1 hour).
///
/// Can be configured via AI_REQUEST_TIMEOUT_SECONDS environment variable.
/// Valid range: 1-86400 seconds (24 hours).
/// - Minimum (1s): Prevents immediate timeout, allows minimal response time
/// - Maximum (24h): Prevents indefinite hangs while supporting long-running AI operations
/// - Default (1h): Balances responsiveness with support for complex AI tasks
///
/// Note: This is a total request timeout, not an idle timeout.
/// Long-running streaming responses that exceed this duration will be terminated,
/// even if actively receiving data.
///
/// CRITICAL: If using a reverse proxy (NGINX, Traefik, etc.), you MUST configure
/// proxy timeouts to match or exceed this value. Without proper proxy configuration,
/// connections will be terminated prematurely at the proxy layer regardless of this
/// backend timeout setting.
///
/// Example NGINX configuration:
/// location /api/ {
/// proxy_read_timeout 3600s; # Must be >= AI_REQUEST_TIMEOUT_SECONDS
/// proxy_send_timeout 3600s;
/// proxy_connect_timeout 60s;
/// }
static ref AI_TIMEOUT_SECS: u64 = {
match std::env::var("AI_REQUEST_TIMEOUT_SECONDS")
.ok()
.and_then(|s| s.parse::<u64>().ok())
{
Some(timeout) if timeout >= AI_TIMEOUT_MIN_SECS && timeout <= AI_TIMEOUT_MAX_SECS => {
tracing::info!("AI request timeout configured: {}s", timeout);
timeout
},
Some(timeout) => {
tracing::warn!(
"AI_REQUEST_TIMEOUT_SECONDS value {} is out of range ({}-{}), using default {}s",
timeout,
AI_TIMEOUT_MIN_SECS,
AI_TIMEOUT_MAX_SECS,
AI_TIMEOUT_DEFAULT_SECS
);
AI_TIMEOUT_DEFAULT_SECS
},
None => {
tracing::info!(
"AI_REQUEST_TIMEOUT_SECONDS not set, using default {}s",
AI_TIMEOUT_DEFAULT_SECS
);
AI_TIMEOUT_DEFAULT_SECS
},
}
};
static ref HTTP_CLIENT: Client = configure_client(reqwest::ClientBuilder::new()
.timeout(std::time::Duration::from_secs(*AI_TIMEOUT_SECS))
.pool_max_idle_per_host(HTTP_POOL_MAX_IDLE_PER_HOST)
.pool_idle_timeout(Some(std::time::Duration::from_secs(HTTP_POOL_IDLE_TIMEOUT_SECS)))
.user_agent("windmill/beta"))
.build()
.expect("Failed to build AI HTTP client - check system TLS configuration");
static ref OPENAI_AZURE_BASE_PATH: Option<String> = std::env::var("OPENAI_AZURE_BASE_PATH").ok();
pub static ref AI_REQUEST_CACHE: Cache<(String, AIProvider), ExpiringAIRequestConfig> = Cache::new(500);
/// Parse AI_HTTP_HEADERS environment variable into a vector of (header_name, header_value) tuples
/// Format: "header1: value1, header2: value2"
static ref AI_HTTP_HEADERS: Vec<(String, String)> = {
std::env::var("AI_HTTP_HEADERS")
.ok()
.map(|headers_str| {
headers_str
.split(',')
.filter_map(|header| {
let parts: Vec<&str> = header.splitn(2, ':').collect();
if parts.len() == 2 {
let name = parts[0].trim().to_string();
let value = parts[1].trim().to_string();
if !name.is_empty() && !value.is_empty() {
Some((name, value))
} else {
None
}
} else {
None
}
})
.collect()
})
.unwrap_or_default()
};
}
#[derive(Deserialize, Debug)]
struct AIOAuthResource {
client_id: String,
client_secret: String,
token_url: String,
user: Option<String>,
}
/// Platform for Anthropic API
#[derive(Deserialize, Debug, Clone, Default, PartialEq)]
#[serde(rename_all = "snake_case")]
enum AnthropicPlatform {
#[default]
Standard,
GoogleVertexAi,
}
#[derive(Deserialize, Debug)]
struct AIStandardResource {
#[serde(alias = "baseUrl", default, deserialize_with = "empty_string_as_none")]
base_url: Option<String>,
#[serde(alias = "apiKey", default, deserialize_with = "empty_string_as_none")]
api_key: Option<String>,
#[serde(default, deserialize_with = "empty_string_as_none")]
organization_id: Option<String>,
#[serde(default, deserialize_with = "empty_string_as_none")]
region: Option<String>,
#[serde(
alias = "awsAccessKeyId",
default,
deserialize_with = "empty_string_as_none"
)]
aws_access_key_id: Option<String>,
#[serde(
alias = "awsSecretAccessKey",
default,
deserialize_with = "empty_string_as_none"
)]
aws_secret_access_key: Option<String>,
/// Platform for Anthropic API (standard or google_vertex_ai)
#[serde(default)]
platform: AnthropicPlatform,
}
#[derive(Deserialize, Debug)]
struct OAuthTokens {
access_token: String,
}
#[derive(Deserialize, Debug)]
#[serde(untagged)]
enum AIResource {
OAuth(AIOAuthResource),
Standard(AIStandardResource),
}
#[derive(Deserialize, Clone, Debug)]
struct AIRequestConfig {
pub base_url: String,
pub api_key: Option<String>,
pub access_token: Option<String>,
pub organization_id: Option<String>,
pub user: Option<String>,
#[allow(dead_code)]
pub region: Option<String>,
#[allow(dead_code)]
pub aws_access_key_id: Option<String>,
#[allow(dead_code)]
pub aws_secret_access_key: Option<String>,
pub platform: AnthropicPlatform,
}
impl AIRequestConfig {
pub async fn new(
provider: &AIProvider,
db: &DB,
w_id: &str,
resource: AIResource,
) -> Result<Self> {
let (
api_key,
access_token,
organization_id,
base_url,
user,
region,
aws_access_key_id,
aws_secret_access_key,
platform,
) = match resource {
AIResource::Standard(resource) => {
let region = resource.region.clone();
let platform = resource.platform.clone();
// Skip get_base_url for Bedrock - it uses SDK directly, not HTTP
let base_url = if matches!(provider, AIProvider::AWSBedrock) {
String::new()
} else {
provider.get_base_url(resource.base_url, db).await?
};
let api_key = if let Some(api_key) = resource.api_key {
Some(get_variable_or_self(api_key, db, w_id).await?)
} else {
None
};
let organization_id = if let Some(organization_id) = resource.organization_id {
Some(get_variable_or_self(organization_id, db, w_id).await?)
} else {
None
};
let aws_access_key_id = if let Some(access_key_id) = resource.aws_access_key_id {
Some(get_variable_or_self(access_key_id, db, w_id).await?)
} else {
None
};
let aws_secret_access_key =
if let Some(secret_access_key) = resource.aws_secret_access_key {
Some(get_variable_or_self(secret_access_key, db, w_id).await?)
} else {
None
};
(
api_key,
None,
organization_id,
base_url,
None,
region,
aws_access_key_id,
aws_secret_access_key,
platform,
)
}
AIResource::OAuth(resource) => {
let user = if let Some(user) = resource.user.clone() {
Some(get_variable_or_self(user, db, w_id).await?)
} else {
None
};
let token = Self::get_token_using_oauth(resource, db, w_id).await?;
let base_url = provider.get_base_url(None, db).await?;
(
None,
Some(token),
None,
base_url,
user,
None,
None,
None,
AnthropicPlatform::Standard,
)
}
};
Ok(Self {
base_url,
organization_id,
api_key,
access_token,
user,
region,
aws_access_key_id,
aws_secret_access_key,
platform,
})
}
async fn get_token_using_oauth(
mut resource: AIOAuthResource,
db: &DB,
w_id: &str,
) -> Result<String> {
resource.client_id = get_variable_or_self(resource.client_id, db, w_id).await?;
resource.client_secret = get_variable_or_self(resource.client_secret, db, w_id).await?;
resource.token_url = get_variable_or_self(resource.token_url, db, w_id).await?;
let mut params = HashMap::new();
params.insert("grant_type", "client_credentials");
params.insert("scope", "https://cognitiveservices.azure.com/.default");
let response = HTTP_CLIENT
.post(resource.token_url)
.form(&params)
.basic_auth(resource.client_id, Some(resource.client_secret))
.send()
.await
.and_then(|r| r.error_for_status())
.map_err(|err| {
Error::internal_err(format!(
"Failed to get access token using credentials flow: {}",
err
))
})?;
let response = response.json::<OAuthTokens>().await.map_err(|err| {
Error::internal_err(format!(
"Failed to parse access token from credentials flow: {}",
err
))
})?;
Ok(response.access_token)
}
pub fn prepare_request(
self,
provider: &AIProvider,
path: &str,
method: Method,
headers: HeaderMap,
body: Bytes,
) -> Result<RequestBuilder> {
let body = if let Some(user) = self.user {
Self::add_user_to_body(body, user)?
} else {
body
};
let base_url = self.base_url.trim_end_matches('/');
let is_azure = provider.is_azure_openai(base_url);
let is_anthropic = matches!(provider, AIProvider::Anthropic);
let is_anthropic_vertex =
is_anthropic && self.platform == AnthropicPlatform::GoogleVertexAi;
let is_anthropic_sdk = headers.get("X-Anthropic-SDK").is_some();
let is_google_ai = matches!(provider, AIProvider::GoogleAI);
// GoogleAI uses OpenAI-compatible endpoint in the proxy (for the chat), but not for the ai agent
let base_url = if is_google_ai {
format!("{}/openai", base_url)
} else {
base_url.to_string()
};
let base_url = base_url.as_str();
// Build URL based on provider
let (url, body) = if is_anthropic_vertex && method != Method::GET {
let (model, transformed_body) = transform_anthropic_for_vertex(&body)?;
let vertex_url = format!("{}/{}:streamRawPredict", base_url, model);
(vertex_url, transformed_body)
} else if is_azure {
let azure_url = AIProvider::build_azure_openai_url(base_url, path);
(azure_url, body)
} else if is_anthropic_sdk {
let truncated_base_url = base_url.trim_end_matches("/v1");
let anthropic_url = format!("{}/{}", truncated_base_url, path);
(anthropic_url, body)
} else {
let default_url = format!("{}/{}", base_url, path);
(default_url, body)
};
tracing::debug!("AI request URL: {}", url);
let mut request = HTTP_CLIENT
.request(method.clone(), &url)
.header("content-type", "application/json");
for (header_name, header_value) in headers.iter() {
// Forward anthropic-* headers, but skip anthropic-version for Vertex AI
// (Vertex AI requires anthropic_version in the request body, not as a header)
if header_name.to_string().starts_with("anthropic-") {
if is_anthropic_vertex && header_name.as_str() == "anthropic-version" {
continue;
}
request = request.header(header_name, header_value);
}
}
// Add authentication headers
if let Some(api_key) = self.api_key {
if is_azure {
request = request.header("api-key", api_key.clone())
} else {
request = request.header("authorization", format!("Bearer {}", api_key.clone()))
}
// For standard Anthropic API, also add X-API-Key header (but not for Vertex AI)
if is_anthropic && !is_anthropic_vertex {
request = request.header("X-API-Key", api_key);
}
}
if let Some(access_token) = self.access_token {
request = request.header("authorization", format!("Bearer {}", access_token))
}
request = request.body(body);
if let Some(org_id) = self.organization_id {
request = request.header("OpenAI-Organization", org_id);
}
// Apply custom headers from AI_HTTP_HEADERS environment variable
for (header_name, header_value) in AI_HTTP_HEADERS.iter() {
request = request.header(header_name.as_str(), header_value.as_str());
}
Ok(request)
}
fn add_user_to_body(body: Bytes, user: String) -> Result<Bytes> {
tracing::debug!("Adding user to request body");
let mut json_body: HashMap<String, Box<RawValue>> = serde_json::from_slice(&body)
.map_err(|e| Error::internal_err(format!("Failed to parse request body: {}", e)))?;
let user_json_string = serde_json::Value::String(user).to_string(); // makes sure to escape characters
json_body.insert(
"user".to_string(),
RawValue::from_string(user_json_string)
.map_err(|e| Error::internal_err(format!("Failed to parse user: {}", e)))?,
);
Ok(serde_json::to_vec(&json_body)
.map_err(|e| Error::internal_err(format!("Failed to reserialize request body: {}", e)))?
.into())
}
}
#[derive(Clone, Debug)]
pub struct ExpiringAIRequestConfig {
config: AIRequestConfig,
expires_at: std::time::Instant,
}
impl ExpiringAIRequestConfig {
fn new(config: AIRequestConfig) -> Self {
Self { config, expires_at: std::time::Instant::now() + std::time::Duration::from_secs(60) }
}
fn is_expired(&self) -> bool {
self.expires_at < std::time::Instant::now()
}
}
#[derive(Serialize, Deserialize, Debug)]
pub struct AIConfig {
#[serde(skip_serializing_if = "Option::is_none")]
pub providers: Option<HashMap<AIProvider, ProviderConfig>>,
#[serde(skip_serializing_if = "Option::is_none")]
pub default_model: Option<ProviderModel>,
#[serde(skip_serializing_if = "Option::is_none")]
pub code_completion_model: Option<ProviderModel>,
#[serde(skip_serializing_if = "Option::is_none")]
pub custom_prompts: Option<HashMap<String, String>>,
#[serde(skip_serializing_if = "Option::is_none")]
pub max_tokens_per_model: Option<HashMap<String, i32>>,
}
/// Anthropic API version for Google Vertex AI
const ANTHROPIC_VERSION_VERTEX: &str = "vertex-2023-10-16";
/// Transforms an Anthropic request for Google Vertex AI:
/// - Extracts the model from the body (needed for the URL)
/// - Adds anthropic_version to the body
fn transform_anthropic_for_vertex(body: &Bytes) -> Result<(String, Bytes)> {
let mut json_body: HashMap<String, serde_json::Value> = serde_json::from_slice(body)
.map_err(|e| Error::internal_err(format!("Failed to parse Anthropic request: {}", e)))?;
// Extract and remove model from body
let model = json_body
.remove("model")
.and_then(|v| v.as_str().map(|s| s.to_string()))
.ok_or_else(|| {
Error::BadRequest("Missing 'model' field in Anthropic request".to_string())
})?;
// Add anthropic_version to body (required for Vertex AI)
json_body.insert(
"anthropic_version".to_string(),
serde_json::Value::String(ANTHROPIC_VERSION_VERTEX.to_string()),
);
let transformed_body = serde_json::to_vec(&json_body)
.map_err(|e| Error::internal_err(format!("Failed to serialize Vertex request: {}", e)))?;
Ok((model, Bytes::from(transformed_body)))
}
// FIM (Fill-in-the-Middle) simulation for providers that don't support native FIM
#[derive(Deserialize, Debug)]
struct FimRequest {
model: String,
prompt: String, // code before cursor
suffix: Option<String>, // code after cursor
temperature: Option<f32>,
max_tokens: Option<u32>,
stop: Option<Vec<String>>,
}
/// Checks if the AI provider supports native FIM (Fill-in-the-Middle) endpoint
fn supports_native_fim(provider: &AIProvider) -> bool {
matches!(provider, AIProvider::Mistral)
}
/// Transforms a FIM request to chat/completions format for providers that don't support native FIM.
fn transform_fim_to_chat_completions(body: &Bytes) -> Result<(Bytes, String)> {
let fim_req: FimRequest = serde_json::from_slice(body)
.map_err(|e| Error::internal_err(format!("Failed to parse FIM request: {}", e)))?;
let suffix = fim_req.suffix.unwrap_or_default();
let system_prompt = "You are a code completion assistant. Complete the code at the <CURSOR/> position between the given prefix and suffix. Output ONLY the code that goes at the cursor - no explanations, no markdown, no repeating the prefix or suffix.";
let user_content = format!(
"<PREFIX>\n{}\n<CURSOR/>\n<SUFFIX>\n{}",
fim_req.prompt, suffix
);
let chat_req = json!({
"model": fim_req.model,
"messages": [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_content}
],
"temperature": fim_req.temperature.unwrap_or(0.0),
"max_tokens": fim_req.max_tokens.unwrap_or(256),
"stop": fim_req.stop
});
let chat_body = serde_json::to_vec(&chat_req)
.map_err(|e| Error::internal_err(format!("Failed to serialize chat request: {}", e)))?;
Ok((Bytes::from(chat_body), "chat/completions".to_string()))
}
pub fn global_service() -> Router {
Router::new().route("/proxy/*ai", post(global_proxy).get(global_proxy))
}
pub fn workspaced_service() -> Router {
let router = Router::new().route("/proxy/*ai", post(proxy).get(proxy));
#[cfg(feature = "bedrock")]
let router = router.route("/check_bedrock_credentials", get(check_bedrock_credentials));
router
}
/// Check if AWS Bedrock credentials are available from environment variables.
#[cfg(feature = "bedrock")]
async fn check_bedrock_credentials(
_authed: ApiAuthed,
Path(_w_id): Path<String>,
) -> Result<Json<windmill_common::ai_bedrock::BedrockCredentialsCheck>> {
let response = windmill_common::ai_bedrock::check_env_credentials().await;
Ok(Json(response))
}
fn is_sse_response(headers: &HeaderMap) -> bool {
headers
.get(http::header::CONTENT_TYPE)
.and_then(|v| v.to_str().ok())
.map(|ct| ct.contains("text/event-stream"))
.unwrap_or(false)
}
fn inject_keepalives<S>(
upstream: S,
interval: Duration,
) -> impl futures::Stream<Item = std::result::Result<Bytes, reqwest::Error>>
where
S: futures::Stream<Item = std::result::Result<Bytes, reqwest::Error>> + Unpin,
{
async_stream::stream! {
tokio::pin!(upstream);
loop {
tokio::select! {
biased;
chunk = upstream.next() => {
match chunk {
Some(item) => yield item,
None => break,
}
}
_ = tokio::time::sleep(interval) => {
yield Ok(Bytes::from(": keepalive\n\n"));
}
}
}
}
}
async fn global_proxy(
authed: ApiAuthed,
Extension(db): Extension<DB>,
Path(ai_path): Path<String>,
method: Method,
headers: HeaderMap,
body: Bytes,
) -> impl IntoResponse {
let provider = headers
.get("X-Provider")
.map(|v| v.to_str().unwrap_or("").to_string());
let api_key = headers
.get("X-API-Key")
.map(|v| v.to_str().unwrap_or("").to_string());
let provider = match provider {
Some(provider) => AIProvider::try_from(provider.as_str())?,
None => return Err(Error::BadRequest("Provider is required".to_string())),
};
let Some(api_key) = api_key else {
return Err(Error::BadRequest("API key is required".to_string()));
};
let base_url = provider.get_base_url(None, &db).await?;
let url = format!("{}/{}", base_url, ai_path);
let mut request = HTTP_CLIENT
.request(method, url)
.header("content-type", "application/json")
.header("Authorization", format!("Bearer {}", api_key));
// Apply custom headers from AI_HTTP_HEADERS environment variable
for (header_name, header_value) in AI_HTTP_HEADERS.iter() {
request = request.header(header_name.as_str(), header_value.as_str());
}
let request = request.body(body);
let response = request.send().await.map_err(to_anyhow)?;
let mut tx = db.begin().await?;
audit_log(
&mut *tx,
&authed,
"ai.global_request",
ActionKind::Execute,
"global",
Some(&authed.email),
None,
)
.await?;
tx.commit().await?;
if response.error_for_status_ref().is_err() {
let err_msg = response.text().await.unwrap_or("".to_string());
return Err(Error::AIError(err_msg));
}
let status_code = response.status();
let headers = response.headers().clone();
let stream = response.bytes_stream();
let body = if is_sse_response(&headers) {
axum::body::Body::from_stream(inject_keepalives(
stream,
Duration::from_secs(KEEPALIVE_INTERVAL_SECS),
))
} else {
axum::body::Body::from_stream(stream)
};
Ok((status_code, headers, body))
}
async fn proxy(
authed: ApiAuthed,
Extension(db): Extension<DB>,
Path((w_id, mut ai_path)): Path<(String, String)>,
method: Method,
headers: HeaderMap,
mut body: Bytes,
) -> impl IntoResponse {
let provider = headers
.get("X-Provider")
.map(|v| v.to_str().unwrap_or("").to_string());
let provider = match provider {
Some(provider) => AIProvider::try_from(provider.as_str())?,
None => return Err(Error::BadRequest("Provider is required".to_string())),
};
let workspace_cache = AI_REQUEST_CACHE.get(&(w_id.clone(), provider.clone()));
let forced_resource_path = headers
.get("X-Resource-Path")
.map(|v| v.to_str().unwrap_or("").to_string());
let request_config = match workspace_cache {
Some(request_cache) if !request_cache.is_expired() && forced_resource_path.is_none() => {
request_cache.config
}
_ => {
let (resource_path, save_to_cache) = if let Some(resource_path) = forced_resource_path {
// forced resource path
(resource_path, false)
} else {
let ai_config = sqlx::query_scalar!(
"SELECT ai_config FROM workspace_settings WHERE workspace_id = $1",
&w_id
)
.fetch_one(&db)
.await?;
if ai_config.is_none() {
return Err(Error::internal_err(
"AI resource not configured".to_string(),
));
}
let mut ai_config = serde_json::from_value::<AIConfig>(ai_config.unwrap())
.map_err(|e| Error::BadRequest(e.to_string()))?;
let provider_config = ai_config
.providers
.as_mut()
.map(|providers| providers.remove(&provider))
.flatten()
.ok_or_else(|| {
Error::BadRequest(format!("Provider {:?} not configured", provider))
})?;
if provider_config.resource_path.is_empty() {
return Err(Error::BadRequest("Resource path is empty".to_string()));
}
(provider_config.resource_path, true)
};
let resource= sqlx::query_scalar!(
"SELECT value as \"value: sqlx::types::Json<Box<RawValue>>\" FROM resource WHERE path = $1 AND workspace_id = $2",
&resource_path,
&w_id
)
.fetch_optional(&db)
.await?
.ok_or_else(|| Error::NotFound(format!("Could not find the resource {}, update the resource path in the workspace settings", resource_path)))?
.ok_or_else(|| Error::BadRequest(format!("Empty resource value for {}", resource_path)))?;
let resource = serde_json::from_str::<AIResource>(resource.0.get())
.map_err(|e| Error::BadRequest(e.to_string()))?;
let request_config = AIRequestConfig::new(&provider, &db, &w_id, resource).await?;
if save_to_cache {
AI_REQUEST_CACHE.insert(
(w_id.clone(), provider.clone()),
ExpiringAIRequestConfig::new(request_config.clone()),
);
}
request_config
}
};
// Check if this is a FIM request to a provider that doesn't support native FIM endpoint
// For such providers, transform to use FIM sentinel tokens with the chat/completions endpoint
let is_fim_request = ai_path.contains("fim/completions");
if is_fim_request && !supports_native_fim(&provider) {
tracing::debug!(
"Transforming FIM request to chat/completions with FIM tokens for provider {:?}",
provider
);
let (chat_body, chat_path) = transform_fim_to_chat_completions(&body)?;
body = chat_body;
ai_path = chat_path;
}
// Handle Bedrock-specific logic when the feature is enabled
#[cfg(feature = "bedrock")]
{
// Extract model and streaming flag for Bedrock transformation (only for POST requests)
let (model, is_streaming) = if matches!(provider, AIProvider::AWSBedrock)
&& method == Method::POST
{
#[derive(Deserialize, Debug)]
struct BedrockRequest {
model: String,
#[serde(default)]
stream: bool,
}
let parsed: BedrockRequest = serde_json::from_slice(&body)
.map_err(|e| Error::internal_err(format!("Failed to parse request body: {}", e)))?;
(Some(parsed.model), parsed.stream)
} else {
(None, false)
};
// For Bedrock requests, use the SDK-based approach
if matches!(provider, AIProvider::AWSBedrock) {
let region = request_config
.region
.as_deref()
.unwrap_or(windmill_common::ai_providers::USE_ENV_REGION);
// Audit log before making the SDK request
let mut tx = db.begin().await?;
audit_log(
&mut *tx,
&authed,
"ai.request",
ActionKind::Execute,
&w_id,
Some(&authed.email),
Some([("ai_config_path", &format!("{:?}", ai_path)[..])].into()),
)
.await?;
tx.commit().await?;
// Handle GET requests for control plane operations
if method == Method::GET {
if ai_path == "foundation-models" {
return bedrock::list_foundation_models(
request_config.api_key.as_deref(),
request_config.aws_access_key_id.as_deref(),
request_config.aws_secret_access_key.as_deref(),
region,
)
.await;
} else if ai_path == "inference-profiles" {
return bedrock::list_inference_profiles(
request_config.api_key.as_deref(),
request_config.aws_access_key_id.as_deref(),
request_config.aws_secret_access_key.as_deref(),
region,
)
.await;
}
}
// Handle POST requests for inference
if method == Method::POST && model.is_some() {
if is_streaming {
return bedrock::handle_bedrock_sdk_streaming(
model.as_ref().unwrap(),
&body,
request_config.api_key.as_deref(),
request_config.aws_access_key_id.as_deref(),
request_config.aws_secret_access_key.as_deref(),
region,
)
.await;
} else {
return bedrock::handle_bedrock_sdk_non_streaming(
model.as_ref().unwrap(),
&body,
request_config.api_key.as_deref(),
request_config.aws_access_key_id.as_deref(),
request_config.aws_secret_access_key.as_deref(),
region,
)
.await;
}
}
}
}
// When bedrock feature is disabled, return error for Bedrock provider
#[cfg(not(feature = "bedrock"))]
if matches!(provider, AIProvider::AWSBedrock) {
return Err(Error::BadRequest(
"AWS Bedrock support is not enabled. Build with 'bedrock' feature.".to_string(),
));
}
let request = request_config.prepare_request(&provider, &ai_path, method, headers, body)?;
let response = request.send().await.map_err(to_anyhow)?;
let mut tx = db.begin().await?;
audit_log(
&mut *tx,
&authed,
"ai.request",
ActionKind::Execute,
&w_id,
Some(&authed.email),
Some([("ai_config_path", &format!("{:?}", ai_path)[..])].into()),
)
.await?;
tx.commit().await?;
if response.error_for_status_ref().is_err() {
let err_msg = response.text().await.unwrap_or("".to_string());
return Err(Error::AIError(err_msg));
}
let status_code = response.status();
let headers = response.headers().clone();
let stream = response.bytes_stream();
let body = if is_sse_response(&headers) {
axum::body::Body::from_stream(inject_keepalives(
stream,
Duration::from_secs(KEEPALIVE_INTERVAL_SECS),
))
} else {
axum::body::Body::from_stream(stream)
};
Ok((status_code, headers, body))
}