jacob/katniss-ai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: novita, nvidia nim, vercel ai gateway

Browse Source
This commit is contained in:
Jacob
2026-10-13 23:37:24 -04:00
parent a0d95d52e4
commit f4f657c219
6 changed files with 918 additions and 794 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/lib.rs
View File

@@ -54,9 +54,18 @@ pub struct AppArgs {
#[arg(long)]
pub mistral_api_key: Vec<SecretString>,
#[arg(long)]
pub novita_ai_api_key: Vec<SecretString>,
#[arg(long)]
pub nvidia_nim_api_key: Vec<SecretString>,
#[arg(long)]
pub openrouter_api_key: Vec<SecretString>,
#[arg(long)]
pub vercel_ai_gateway_api_key: Vec<SecretString>,
#[arg(long)]
pub z_ai_api_key: Vec<SecretString>,
}

1110
src/server/chat/completions.rs
View File

File diff suppressed because it is too large Load Diff

9
src/server/decisions.rs
View File

@@ -224,8 +224,13 @@ pub async fn create_decision(
.into_iter()
.map(|(choice_id, probability)| {
openai::decisions::ScoreProbability {
value: choice_id.parse().expect("it should be a number, I think"),
label: legend.get(&choice_id).expect("it should exist, I think").clone(),
value: choice_id
.parse()
.expect("it should be a number, I think"),
label: legend
.get(&choice_id)
.expect("it should exist, I think")
.clone(),
probability,
}
})

565
src/server/embeddings.rs
View File

@@ -1,279 +1,286 @@
use crate::CloudflareCredentials;
use crate::openai;
use crate::server::AppState;
use crate::server::Provider;
use crate::server::Upstream;
use crate::server::take_random_by_swapping;
use axum::Json;
use axum::extract::State;
use axum::http::StatusCode;
use axum::response::IntoResponse;
use axum::response::Response;
use axum_extra::routing::TypedPath;
use ecow::EcoString;
use headers::Header;
use headers::authorization::Bearer;
use mitsein::vec1::Vec1;
use reqwest::Certificate;
use secrecy::ExposeSecret;
use serde::Deserialize;
use serde_with::DeserializeFromStr;
use serde_with::SerializeDisplay;
use std::collections::VecDeque;
use strum::EnumString;
use strum::VariantArray;
#[derive(Deserialize, TypedPath)]
#[typed_path("/v1/embeddings")]
pub struct V1EmbeddingsPath;
#[derive(
Clone, Debug, EnumString, DeserializeFromStr, strum::Display, SerializeDisplay, VariantArray,
)]
#[strum(serialize_all = "lowercase")]
pub enum Shortcut {
#[strum(serialize = "codestral-embed-2505")]
CodestralEmbed2505,
#[strum(serialize = "lfm2.5-embedding-350m")]
Lfm2_5Embedding350M,
#[strum(serialize = "mistral-embed-2312")]
MistralEmbed2312,
#[strum(serialize = "nemotron-3-embed-1b")]
Nemotron3Embed1B,
}
impl Shortcut {
fn suggestions(&self) -> Vec<(Provider, EcoString)> {
use Provider as P;
use Shortcut as S;
match self {
S::CodestralEmbed2505 => vec![
(P::Mistral, "codestral-embed-2505".into())
],
S::Lfm2_5Embedding350M => {
vec![(P::OpenRouter, "liquid/lfm-2.5-embedding-350m:free".into())]
},
S::MistralEmbed2312 => vec![
(P::Mistral, "mistral-embed-2312".into())
],
S::Nemotron3Embed1B => {
vec![(P::OpenRouter, "nvidia/nemotron-3-embed-1b:free".into())]
}
}
}
}
pub struct EmbeddingsResponse(openai::embeddings::CreateEmbeddingResponse<String>);
impl IntoResponse for EmbeddingsResponse {
fn into_response(self) -> Response {
let Self(embeddings) = self;
Json(embeddings).into_response()
}
}
pub enum EmbeddingsError {
NoneSucceeded,
}
impl IntoResponse for EmbeddingsError {
fn into_response(self) -> Response {
match self {
EmbeddingsError::NoneSucceeded => StatusCode::SERVICE_UNAVAILABLE.into_response(),
}
}
}
#[tracing::instrument(skip(state, request))]
pub async fn create_embeddings(
_: V1EmbeddingsPath,
State(state): State<AppState>,
Json(request): Json<openai::embeddings::CreateEmbeddingRequest<Shortcut>>,
) -> Result<EmbeddingsResponse, EmbeddingsError> {
let shortcut = request.model.clone();
let mut model_upstreams = VecDeque::new();
for (provider, model) in shortcut.suggestions() {
let upstreams: Vec<Upstream> =
match provider {
Provider::CloudflareWorkersAi => state
.app_args
.cloudflare_workers_ai_credentials
.iter()
.cloned()
.map(
|CloudflareCredentials {
account_id,
api_key,
}| {
Upstream::builder()
.url(
format!("https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1/embeddings")
.parse()
.expect("TODO"),
)
.api_key(api_key)
.build()
},
)
.collect(),
Provider::GoogleAiStudio => state
.app_args
.google_ai_studio_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url(
"https://generativelanguage.googleapis.com/v1beta/openai/v1/embeddings"
.parse()
.expect("TODO"),
)
.api_key(api_key)
.build()
})
.collect(),
Provider::Groq => state
.app_args
.groq_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://api.groq.com/openai/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::HuggingFace => state
.app_args
.hugging_face_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://router.huggingface.co/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::LiquidAi => unreachable!(),
Provider::Mistral => state
.app_args
.mistral_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://api.mistral.ai/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::OpenRouter => state
.app_args
.openrouter_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://openrouter.ai/api/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::ZAi => unreachable!(),
};
if let Ok(upstreams) = Vec1::try_from(upstreams) {
model_upstreams.push_back((model, upstreams));
}
}
while let Some((model, upstreams)) = model_upstreams.pop_front() {
let (upstream, upstreams) = take_random_by_swapping(upstreams, &mut rand::rng());
if let Some(upstreams) = upstreams {
model_upstreams.push_back((model.clone(), upstreams));
}
tracing::info!(?model, ?upstream);
let openai::embeddings::CreateEmbeddingRequest {
model: _shortcut,
input,
dimensions,
encoding_format,
user,
} = request.clone();
let request = openai::embeddings::CreateEmbeddingRequest {
model,
input,
dimensions,
encoding_format,
user,
};
let client = reqwest::Client::builder()
.tls_certs_only(
webpki_root_certs::TLS_SERVER_ROOT_CERTS
.iter()
.filter_map(|cert| Certificate::from_der(cert).ok()),
)
.build()
.expect("TODO");
let mut upstream_request = client
.post(upstream.url)
.json(&request);
if let Some(api_key) = upstream.api_key {
let authorization =
headers::Authorization::bearer(api_key.expose_secret()).expect("TODO");
upstream_request = upstream_request.header(headers::Authorization::<Bearer>::name(), {
let mut vec = vec![];
authorization.encode(&mut vec);
let [one] = vec.try_into().expect("TODO");
one
});
}
let upstream_response = match upstream_request.send().await {
Ok(upstream_response) => upstream_response,
Err(upstream_error) => {
tracing::error!(?upstream_error);
continue;
}
};
if let Err(upstream_error) = upstream_response.error_for_status_ref() {
tracing::error!(?upstream_error);
match upstream_response.text().await {
Ok(body) => tracing::error!(?body),
Err(body_error) => tracing::error!(?body_error),
};
continue;
};
let response = match upstream_response.json().await {
Ok(response) => response,
Err(json_error) => {
tracing::error!(?json_error);
continue;
}
};
return Ok(EmbeddingsResponse(response));
}
Err(EmbeddingsError::NoneSucceeded)
}
use crate::CloudflareCredentials;
use crate::openai;
use crate::server::AppState;
use crate::server::Provider;
use crate::server::Upstream;
use crate::server::take_random_by_swapping;
use axum::Json;
use axum::extract::State;
use axum::http::StatusCode;
use axum::response::IntoResponse;
use axum::response::Response;
use axum_extra::routing::TypedPath;
use ecow::EcoString;
use headers::Header;
use headers::authorization::Bearer;
use mitsein::vec1::Vec1;
use reqwest::Certificate;
use secrecy::ExposeSecret;
use serde::Deserialize;
use serde_with::DeserializeFromStr;
use serde_with::SerializeDisplay;
use std::collections::VecDeque;
use strum::EnumString;
use strum::VariantArray;
#[derive(Deserialize, TypedPath)]
#[typed_path("/v1/embeddings")]
pub struct V1EmbeddingsPath;
#[derive(
Clone, Debug, EnumString, DeserializeFromStr, strum::Display, SerializeDisplay, VariantArray,
)]
#[strum(serialize_all = "lowercase")]
pub enum Shortcut {
#[strum(serialize = "codestral-embed-2505")]
CodestralEmbed2505,
#[strum(serialize = "lfm2.5-embedding-350m")]
Lfm2_5Embedding350M,
#[strum(serialize = "mistral-embed-2312")]
MistralEmbed2312,
#[strum(serialize = "nemotron-3-embed-1b")]
Nemotron3Embed1B,
}
impl Shortcut {
fn suggestions(&self) -> Vec<(Provider, EcoString)> {
use Provider as P;
use Shortcut as S;
match self {
S::CodestralEmbed2505 => vec![(P::Mistral, "codestral-embed-2505".into())],
S::Lfm2_5Embedding350M => {
vec![(P::OpenRouter, "liquid/lfm-2.5-embedding-350m:free".into())]
}
S::MistralEmbed2312 => vec![(P::Mistral, "mistral-embed-2312".into())],
S::Nemotron3Embed1B => vec![
(P::NvidiaNim, "nvidia/nemotron-3-embed-1b".into()),
(P::OpenRouter, "nvidia/nemotron-3-embed-1b:free".into()),
],
}
}
}
pub struct EmbeddingsResponse(openai::embeddings::CreateEmbeddingResponse<String>);
impl IntoResponse for EmbeddingsResponse {
fn into_response(self) -> Response {
let Self(embeddings) = self;
Json(embeddings).into_response()
}
}
pub enum EmbeddingsError {
NoneSucceeded,
}
impl IntoResponse for EmbeddingsError {
fn into_response(self) -> Response {
match self {
EmbeddingsError::NoneSucceeded => StatusCode::SERVICE_UNAVAILABLE.into_response(),
}
}
}
#[tracing::instrument(skip(state, request))]
pub async fn create_embeddings(
_: V1EmbeddingsPath,
State(state): State<AppState>,
Json(request): Json<openai::embeddings::CreateEmbeddingRequest<Shortcut>>,
) -> Result<EmbeddingsResponse, EmbeddingsError> {
let shortcut = request.model.clone();
let mut model_upstreams = VecDeque::new();
for (provider, model) in shortcut.suggestions() {
let upstreams: Vec<Upstream> =
match provider {
Provider::CloudflareWorkersAi => state
.app_args
.cloudflare_workers_ai_credentials
.iter()
.cloned()
.map(
|CloudflareCredentials {
account_id,
api_key,
}| {
Upstream::builder()
.url(
format!("https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1/embeddings")
.parse()
.expect("TODO"),
)
.api_key(api_key)
.build()
},
)
.collect(),
Provider::GoogleAiStudio => state
.app_args
.google_ai_studio_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url(
"https://generativelanguage.googleapis.com/v1beta/openai/v1/embeddings"
.parse()
.expect("TODO"),
)
.api_key(api_key)
.build()
})
.collect(),
Provider::Groq => state
.app_args
.groq_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://api.groq.com/openai/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::HuggingFace => state
.app_args
.hugging_face_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://router.huggingface.co/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::LiquidAi => unreachable!(),
Provider::Mistral => state
.app_args
.mistral_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://api.mistral.ai/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::NvidiaNim => state
.app_args
.nvidia_nim_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://integrate.api.nvidia.com/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::NovitaAi => unreachable!(),
Provider::OpenRouter => state
.app_args
.openrouter_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://openrouter.ai/api/v1/embeddings".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::VercelAiGateway => unreachable!(),
Provider::ZAi => unreachable!(),
};
if let Ok(upstreams) = Vec1::try_from(upstreams) {
model_upstreams.push_back((model, upstreams));
}
}
while let Some((model, upstreams)) = model_upstreams.pop_front() {
let (upstream, upstreams) = take_random_by_swapping(upstreams, &mut rand::rng());
if let Some(upstreams) = upstreams {
model_upstreams.push_back((model.clone(), upstreams));
}
tracing::info!(?model, ?upstream);
let openai::embeddings::CreateEmbeddingRequest {
model: _shortcut,
input,
dimensions,
encoding_format,
user,
} = request.clone();
let request = openai::embeddings::CreateEmbeddingRequest {
model,
input,
dimensions,
encoding_format,
user,
};
let client = reqwest::Client::builder()
.tls_certs_only(
webpki_root_certs::TLS_SERVER_ROOT_CERTS
.iter()
.filter_map(|cert| Certificate::from_der(cert).ok()),
)
.build()
.expect("TODO");
let mut upstream_request = client.post(upstream.url).json(&request);
if let Some(api_key) = upstream.api_key {
let authorization =
headers::Authorization::bearer(api_key.expose_secret()).expect("TODO");
upstream_request = upstream_request.header(headers::Authorization::<Bearer>::name(), {
let mut vec = vec![];
authorization.encode(&mut vec);
let [one] = vec.try_into().expect("TODO");
one
});
}
let upstream_response = match upstream_request.send().await {
Ok(upstream_response) => upstream_response,
Err(upstream_error) => {
tracing::error!(?upstream_error);
continue;
}
};
if let Err(upstream_error) = upstream_response.error_for_status_ref() {
tracing::error!(?upstream_error);
match upstream_response.text().await {
Ok(body) => tracing::error!(?body),
Err(body_error) => tracing::error!(?body_error),
};
continue;
};
let response = match upstream_response.json().await {
Ok(response) => response,
Err(json_error) => {
tracing::error!(?json_error);
continue;
}
};
return Ok(EmbeddingsResponse(response));
}
Err(EmbeddingsError::NoneSucceeded)
}

3
src/server/mod.rs
View File

@@ -28,7 +28,10 @@ enum Provider {
HuggingFace,
LiquidAi,
Mistral,
NovitaAi,
NvidiaNim,
OpenRouter,
VercelAiGateway,
ZAi,
}

16
src/server/system_one.rs
View File

@@ -49,6 +49,8 @@ impl Shortcut {
(P::CloudflareWorkersAi, "clef-flash".into()),
(P::OpenRouter, "respan/span-01-lite:free".into()),
(P::OpenRouter, "respan/span-01-lite".into()),
(P::VercelAiGateway, "convaiinnovations/laya-free".into()),
(P::VercelAiGateway, "convaiinnovations/laya".into()),
]
}
S::M => {
@@ -132,6 +134,8 @@ pub async fn create_system_one(
})
.collect(),
Provider::Mistral => unreachable!(),
Provider::NovitaAi => unreachable!(),
Provider::NvidiaNim => unreachable!(),
Provider::OpenRouter => state
.app_args
.openrouter_api_key
@@ -144,6 +148,18 @@ pub async fn create_system_one(
.build()
})
.collect(),
Provider::VercelAiGateway => state
.app_args
.vercel_ai_gateway_api_key
.iter()
.cloned()
.map(|api_key| {
Upstream::builder()
.url("https://ai-gateway.vercel.sh/typesafe/v1/systemone".parse().expect("TODO"))
.api_key(api_key)
.build()
})
.collect(),
Provider::ZAi => unreachable!(),
};