gitdataai/libs/service/agent/sync.rs

754 lines
25 KiB
Rust

#![allow(dead_code)]
//! Synchronizes AI model metadata from the upstream AI endpoint
//! (`GET /v1/models`) into the local database.
//!
//! Flow:
//! 1. Call `GET /v1/models` with the configured AI API key.
//! 2. Parse the rich response (name, context_length, max_output_tokens,
//! capabilities, pricing, owned_by) — no external metadata source needed.
//! 3. Upsert provider / model / version / pricing / capability / profile
//! records for all accessible models.
//!
//! Usage: call `start_sync_task()` to launch a background task that syncs
//! immediately and then every 10 minutes. On app startup, run it once
//! eagerly before accepting traffic.
use std::time::Duration;
use tokio::task::JoinHandle;
use tokio::time::interval;
use crate::AppService;
use crate::error::AppError;
use chrono::Utc;
use db::database::AppDatabase;
use models::agents::model::Entity as ModelEntity;
use models::agents::model_capability::Entity as CapabilityEntity;
use models::agents::model_parameter_profile::Entity as ProfileEntity;
use models::agents::model_provider::Entity as ProviderEntity;
use models::agents::model_provider::Model as ProviderModel;
use models::agents::model_version::Entity as VersionEntity;
use models::agents::{CapabilityType, ModelCapability, ModelModality, ModelStatus};
use sea_orm::Set;
use sea_orm::prelude::*;
use serde::Deserialize;
use serde::Serialize;
use session::Session;
use utoipa::ToSchema;
use uuid::Uuid;
// Upstream /v1/models response types -----------------------------------------
#[derive(Debug, Clone, Deserialize)]
struct ModelsListResponse {
data: Vec<UpstreamModel>,
}
#[derive(Debug, Clone, Deserialize)]
struct UpstreamModel {
id: String,
#[serde(default)]
name: Option<String>,
#[serde(default)]
owned_by: Option<String>,
#[serde(default)]
context_length: Option<u64>,
#[serde(default)]
max_output_tokens: Option<u64>,
#[serde(default)]
capabilities: Option<UpstreamCapabilities>,
#[serde(default)]
pricing: Option<UpstreamPricing>,
#[serde(default)]
r#type: Option<String>,
}
#[derive(Debug, Clone, Deserialize)]
struct UpstreamCapabilities {
#[serde(default)]
vision: Option<bool>,
#[serde(default)]
tool_call: Option<bool>,
#[serde(default)]
reasoning: Option<bool>,
}
#[derive(Debug, Clone, Deserialize)]
struct UpstreamPricing {
#[serde(default)]
input: Option<f64>,
#[serde(default)]
output: Option<f64>,
#[serde(default)]
cache_read: Option<f64>,
#[serde(default)]
cache_write: Option<f64>,
#[serde(default)]
unit: Option<String>,
#[serde(default)]
currency: Option<String>,
}
// Response type --------------------------------------------------------------
#[derive(Debug, Clone, Serialize, ToSchema)]
pub struct SyncModelsResponse {
pub models_created: i64,
pub models_updated: i64,
pub models_offline: i64,
pub models_deactivated: i64,
pub versions_created: i64,
pub pricing_created: i64,
pub capabilities_created: i64,
pub profiles_created: i64,
}
// Mapping helpers ------------------------------------------------------------
fn infer_modality(model: &UpstreamModel) -> ModelModality {
if let Some(caps) = &model.capabilities {
if caps.vision == Some(true) {
return ModelModality::Multimodal;
}
}
let lower = model.id.to_lowercase();
if lower.contains("vision")
|| lower.contains("dall-e")
|| lower.contains("gpt-image")
|| lower.contains("gpt-4o")
{
ModelModality::Multimodal
} else if lower.contains("embedding") {
ModelModality::Text
} else if lower.contains("whisper") || lower.contains("audio") {
ModelModality::Audio
} else {
ModelModality::Text
}
}
fn infer_capability(model: &UpstreamModel) -> ModelCapability {
let lower = model.id.to_lowercase();
if lower.contains("embedding") {
ModelCapability::Embedding
} else {
ModelCapability::Chat
}
}
fn context_length(model: &UpstreamModel) -> i64 {
model.context_length.map(|c| c as i64).unwrap_or(8_192)
}
fn max_output_tokens(model: &UpstreamModel) -> Option<i64> {
model.max_output_tokens.map(|v| v as i64)
}
fn capability_list(model: &UpstreamModel) -> Vec<(CapabilityType, bool)> {
let mut caps = Vec::new();
// Function call / tool use
if let Some(u) = &model.capabilities {
if u.tool_call == Some(true) {
caps.push((CapabilityType::ToolUse, true));
}
if u.vision == Some(true) {
caps.push((CapabilityType::Vision, true));
}
}
// Always mark function call as supported by default for chat models
if caps.is_empty() {
caps.push((CapabilityType::FunctionCall, true));
}
caps
}
// Provider helpers -----------------------------------------------------------
fn extract_provider_name(model: &UpstreamModel) -> String {
if let Some(owned) = &model.owned_by {
if !owned.is_empty() {
return normalize_provider_name(owned);
}
}
normalize_provider_name(model.id.split('/').next().unwrap_or("unknown"))
}
fn normalize_provider_name(slug: &str) -> String {
match slug {
"openai" => "openai".to_string(),
"anthropic" => "anthropic".to_string(),
"google" | "google-ai" => "google".to_string(),
"mistralai" => "mistral".to_string(),
"meta-llama" | "meta" => "meta".to_string(),
"deepseek" => "deepseek".to_string(),
"azure" | "azure-openai" => "azure".to_string(),
"x-ai" | "xai" => "xai".to_string(),
"moonshot" => "moonshot".to_string(),
"zai" => "zai".to_string(),
"minimax" => "minimax".to_string(),
"alibaba" | "qwen" => "qwen".to_string(),
s => s.to_string(),
}
}
fn provider_display_name(name: &str) -> String {
match name {
"openai" => "OpenAI".to_string(),
"anthropic" => "Anthropic".to_string(),
"google" => "Google DeepMind".to_string(),
"mistral" => "Mistral AI".to_string(),
"meta" => "Meta".to_string(),
"deepseek" => "DeepSeek".to_string(),
"azure" => "Microsoft Azure".to_string(),
"xai" => "xAI".to_string(),
"moonshot" => "Moonshot AI".to_string(),
"zai" => "Zhipu AI".to_string(),
"minimax" => "MiniMax".to_string(),
"qwen" => "Alibaba Qwen".to_string(),
s => s.to_string(),
}
}
// Upsert helpers -------------------------------------------------------------
async fn upsert_provider(db: &AppDatabase, slug: &str) -> Result<ProviderModel, AppError> {
let display = provider_display_name(slug);
let now = Utc::now();
use models::agents::model_provider::Column as PCol;
if let Some(existing) = ProviderEntity::find()
.filter(PCol::Name.eq(slug))
.one(db)
.await?
{
let mut active: models::agents::model_provider::ActiveModel = existing.into();
active.updated_at = Set(now);
active.update(db).await.map_err(AppError::from)
} else {
let active = models::agents::model_provider::ActiveModel {
id: Set(Uuid::now_v7()),
name: Set(slug.to_string()),
display_name: Set(display.to_string()),
website: Set(None),
status: Set(ModelStatus::Active.to_string()),
created_at: Set(now),
updated_at: Set(now),
};
active.insert(db).await.map_err(AppError::from)
}
}
/// Upserts a model by upstream ID as the deduplication key.
/// Each upstream model ID maps to exactly one row in the local `ai_model` table.
async fn upsert_model_by_id(
db: &AppDatabase,
provider_id: Uuid,
model: &UpstreamModel,
) -> Result<(models::agents::model::Model, bool), AppError> {
let now = Utc::now();
let modality = infer_modality(model);
let capability = infer_capability(model);
let ctx = context_length(model);
let max_out = max_output_tokens(model);
let model_id_str = extract_model_name(model);
use models::agents::model::Column as MCol;
if let Some(existing) = ModelEntity::find()
.filter(MCol::Name.eq(&model_id_str))
.one(db)
.await?
{
// Update existing model (deduplicated by name)
let mut active: models::agents::model::ActiveModel = existing.clone().into();
// Update provider if changed (first provider wins)
active.provider_id = Set(provider_id);
active.context_length = Set(ctx);
active.max_output_tokens = Set(max_out);
active.status = Set(ModelStatus::Active.to_string());
active.updated_at = Set(now);
let updated = active.update(db).await?;
Ok((updated, false))
} else {
// Create new model
let active = models::agents::model::ActiveModel {
id: Set(Uuid::now_v7()),
provider_id: Set(provider_id),
name: Set(model_id_str),
modality: Set(modality.to_string()),
capability: Set(capability.to_string()),
context_length: Set(ctx),
max_output_tokens: Set(max_out),
training_cutoff: Set(None),
is_open_source: Set(false),
status: Set(ModelStatus::Active.to_string()),
created_at: Set(now),
updated_at: Set(now),
..Default::default()
};
let inserted = active.insert(db).await.map_err(AppError::from)?;
Ok((inserted, true))
}
}
async fn upsert_version(
db: &AppDatabase,
model_uuid: Uuid,
) -> Result<(models::agents::model_version::Model, bool), AppError> {
use models::agents::model_version::Column as VCol;
let now = Utc::now();
if let Some(existing) = VersionEntity::find()
.filter(VCol::ModelId.eq(model_uuid))
.filter(VCol::IsDefault.eq(true))
.one(db)
.await?
{
Ok((existing, false))
} else {
let active = models::agents::model_version::ActiveModel {
id: Set(Uuid::now_v7()),
model_id: Set(model_uuid),
version: Set("1".to_string()),
release_date: Set(None),
change_log: Set(None),
is_default: Set(true),
status: Set(ModelStatus::Active.to_string()),
created_at: Set(now),
};
let inserted = active.insert(db).await.map_err(AppError::from)?;
Ok((inserted, true))
}
}
async fn upsert_pricing(
db: &AppDatabase,
version_uuid: Uuid,
pricing: Option<&UpstreamPricing>,
) -> Result<bool, AppError> {
use models::agents::model_pricing::Column as PCol;
use models::agents::model_pricing::Entity as PricingEntity;
let existing = PricingEntity::find()
.filter(PCol::ModelVersionId.eq(version_uuid))
.one(db)
.await?;
if existing.is_some() {
return Ok(false);
}
let (input_price, output_price) = if let Some(p) = pricing {
(
format!("{:.2}", p.input.unwrap_or(0.0)),
format!("{:.2}", p.output.unwrap_or(0.0)),
)
} else {
("0.00".to_string(), "0.00".to_string())
};
let currency = pricing
.and_then(|p| p.currency.clone())
.unwrap_or_else(|| "USD".to_string());
let active = models::agents::model_pricing::ActiveModel {
id: Set(Utc::now().timestamp_millis()),
model_version_id: Set(version_uuid),
input_price_per_1k_tokens: Set(input_price),
output_price_per_1k_tokens: Set(output_price),
currency: Set(currency),
effective_from: Set(Utc::now()),
};
active.insert(db).await.map_err(AppError::from)?;
Ok(true)
}
async fn upsert_capabilities(
db: &AppDatabase,
version_uuid: Uuid,
model: &UpstreamModel,
) -> Result<i64, AppError> {
use models::agents::model_capability::Column as CCol;
let caps = capability_list(model);
let now = Utc::now();
let mut created = 0i64;
for (cap_type, supported) in caps {
let exists = CapabilityEntity::find()
.filter(CCol::ModelVersionId.eq(version_uuid))
.filter(CCol::Capability.eq(cap_type.to_string()))
.one(db)
.await?;
if exists.is_some() {
continue;
}
let active = models::agents::model_capability::ActiveModel {
// FIXME: i64 primary key loses entropy from UUID. Use UUID type in schema.
id: Set(Uuid::now_v7().as_u128() as i64),
// FIXME: version_uuid truncated from 128-bit UUID to i64.
// Schema must be migrated: ALTER COLUMN model_version_id TO UUID type.
model_version_id: Set(version_uuid.as_u128() as i64),
capability: Set(cap_type.to_string()),
is_supported: Set(supported),
created_at: Set(now),
};
active.insert(db).await.map_err(AppError::from)?;
created += 1;
}
Ok(created)
}
async fn upsert_parameter_profile(
db: &AppDatabase,
version_uuid: Uuid,
model: &UpstreamModel,
) -> Result<bool, AppError> {
use models::agents::model_parameter_profile::Column as PCol;
let existing = ProfileEntity::find()
.filter(PCol::ModelVersionId.eq(version_uuid))
.one(db)
.await?;
if existing.is_some() {
return Ok(false);
}
let lower = model.id.to_lowercase();
let (t_min, t_max) = if lower.contains("o1") || lower.contains("o3") {
(1.0, 1.0)
} else {
(0.0, 2.0)
};
let active = models::agents::model_parameter_profile::ActiveModel {
id: Set(Utc::now().timestamp_millis()),
model_version_id: Set(version_uuid),
temperature_min: Set(t_min),
temperature_max: Set(t_max),
top_p_min: Set(0.0),
top_p_max: Set(1.0),
frequency_penalty_supported: Set(true),
presence_penalty_supported: Set(true),
};
active.insert(db).await.map_err(AppError::from)?;
Ok(true)
}
// Core sync logic ------------------------------------------------------------
/// Extracts the API model identifier from an upstream model.
/// Uses the upstream `id` field directly (e.g. "kimi-k2.6") as the model name
/// stored in the database, since this is what AI API calls use as the `model` parameter.
fn extract_model_name(model: &UpstreamModel) -> String {
model.id.clone()
}
/// Deduplicates existing models in the database by name.
/// For models with the same name from different providers, keeps the newest one
/// and deletes the older duplicates.
async fn mark_all_models_offline(db: &AppDatabase) -> Result<i64, AppError> {
use models::agents::model::Column as MCol;
use models::agents::model::Entity as MEntity;
let now = Utc::now();
let updated = MEntity::update_many()
.set(models::agents::model::ActiveModel {
status: Set(ModelStatus::Offline.to_string()),
updated_at: Set(now),
..Default::default()
})
.filter(MCol::Status.eq(ModelStatus::Active.to_string()))
.exec(db.writer())
.await
.map_err(|e| AppError::DatabaseError(e.to_string()))?;
Ok(updated.rows_affected as i64)
}
async fn sync_models_from_upstream(
db: &AppDatabase,
upstream_models: Vec<UpstreamModel>,
) -> SyncModelsResponse {
// Step 1: Mark all existing models as offline
let models_offline = mark_all_models_offline(db).await.unwrap_or(0);
tracing::info!(
upstream_total = upstream_models.len(),
"sync_models_from_upstream: syncing models"
);
let mut models_created = 0i64;
let mut models_updated = 0i64;
let models_deactivated: i64;
let mut versions_created = 0i64;
let mut pricing_created = 0i64;
let mut capabilities_created = 0i64;
let mut profiles_created = 0i64;
for model in &upstream_models {
let provider_slug = extract_provider_name(model);
let provider = match upsert_provider(db, &provider_slug).await {
Ok(p) => p,
Err(e) => {
tracing::warn!(
provider = %provider_slug,
error = ?e,
"sync_models_from_upstream: upsert_provider error"
);
continue;
}
};
let (model_record, _is_new) = match upsert_model_by_id(db, provider.id, model).await {
Ok((m, created)) => {
if created {
models_created += 1;
} else {
models_updated += 1;
}
(m, created)
}
Err(e) => {
tracing::warn!(
model = %model.id,
error = ?e,
"sync_models_from_upstream: upsert_model error"
);
continue;
}
};
let (version_record, version_is_new) = match upsert_version(db, model_record.id).await {
Ok(v) => v,
Err(e) => {
tracing::warn!(
model = %model.id,
error = ?e,
"sync_models_from_upstream: upsert_version error"
);
continue;
}
};
if version_is_new {
versions_created += 1;
}
if upsert_pricing(db, version_record.id, model.pricing.as_ref())
.await
.unwrap_or(false)
{
pricing_created += 1;
}
capabilities_created += upsert_capabilities(db, version_record.id, model)
.await
.unwrap_or(0);
if upsert_parameter_profile(db, version_record.id, model)
.await
.unwrap_or(false)
{
profiles_created += 1;
}
}
// Step 3: Deactivate models that were offline before but exist locally
// (These are models that were added manually and are no longer in sync)
let deactivated = deactivate_orphaned_models(db).await.unwrap_or(0);
models_deactivated = deactivated;
SyncModelsResponse {
models_created,
models_updated,
models_offline,
models_deactivated,
versions_created,
pricing_created,
capabilities_created,
profiles_created,
}
}
/// Deactivates models that were previously marked offline and are not in any active sync.
/// These are manually added models that are no longer needed.
async fn deactivate_orphaned_models(db: &AppDatabase) -> Result<i64, AppError> {
use models::agents::model::Column as MCol;
use models::agents::model::Entity as MEntity;
let now = Utc::now();
let updated = MEntity::update_many()
.set(models::agents::model::ActiveModel {
status: Set(ModelStatus::Deprecated.to_string()),
updated_at: Set(now),
..Default::default()
})
.filter(MCol::Status.eq(ModelStatus::Offline.to_string()))
.exec(db.writer())
.await
.map_err(|e| AppError::DatabaseError(e.to_string()))?;
Ok(updated.rows_affected as i64)
}
// HTTP helpers ---------------------------------------------------------------
/// List models from the upstream AI endpoint (`GET /v1/models`).
async fn list_upstream_models(
client: &reqwest::Client,
base_url: &str,
api_key: &str,
) -> Result<Vec<UpstreamModel>, AppError> {
let base = base_url.trim_end_matches('/');
let url = if base.ends_with("/v1") {
format!("{}/models", base)
} else {
format!("{}/v1/models", base)
};
let resp = client
.get(&url)
.header("Authorization", format!("Bearer {}", api_key))
.send()
.await
.map_err(|e| AppError::InternalServerError(format!("failed to list models: {}", e)))?;
let body = resp
.text()
.await
.map_err(|e| AppError::InternalServerError(format!("failed to read models body: {}", e)))?;
// Try standard OpenAI-compatible format: { "data": [{...}, ...] }
if let Ok(parsed) = serde_json::from_str::<ModelsListResponse>(&body) {
return Ok(parsed.data);
}
// Try raw array: [{...}, ...]
if let Ok(parsed) = serde_json::from_str::<Vec<UpstreamModel>>(&body) {
return Ok(parsed);
}
tracing::warn!(
body = %body.chars().take(500).collect::<String>(),
"list_upstream_models: unknown response format"
);
Err(AppError::InternalServerError(format!(
"unexpected /v1/models response format (first 200 chars): {}",
body.chars().take(200).collect::<String>()
)))
}
fn build_ai_client(
config: &config::AppConfig,
) -> Result<(reqwest::Client, String, String), AppError> {
let api_key = config
.ai_api_key()
.map_err(|e| AppError::InternalServerError(format!("AI API key not configured: {}", e)))?;
let base_url = config
.ai_basic_url()
.unwrap_or_else(|_| "https://api.openai.com".into());
Ok((reqwest::Client::new(), base_url, api_key))
}
fn build_ai_client_from_parts(
api_key: Option<String>,
base_url: Option<String>,
) -> Result<(reqwest::Client, String, String), String> {
let api_key = api_key.ok_or_else(|| "AI API key not configured".to_string())?;
let base_url = base_url.unwrap_or_else(|| "https://api.openai.com".into());
Ok((reqwest::Client::new(), base_url, api_key))
}
// Public API -----------------------------------------------------------------
impl AppService {
/// Sync model metadata from the upstream AI endpoint (`GET /v1/models`).
///
/// Parses the full response (name, context_length, max_output_tokens,
/// capabilities, pricing, owned_by) and upserts all related records.
pub async fn sync_upstream_models(
&self,
_ctx: &Session,
) -> Result<SyncModelsResponse, AppError> {
let (http_client, base_url, api_key) = build_ai_client(&self.config)?;
let upstream_models = list_upstream_models(&http_client, &base_url, &api_key).await?;
tracing::info!(
model_count = upstream_models.len(),
"sync_upstream_models: {} models from upstream endpoint",
upstream_models.len()
);
let result = sync_models_from_upstream(&self.db, upstream_models).await;
tracing::info!(
models_created = result.models_created,
models_updated = result.models_updated,
versions_created = result.versions_created,
pricing_created = result.pricing_created,
capabilities_created = result.capabilities_created,
profiles_created = result.profiles_created,
"sync_upstream_models: complete"
);
Ok(result)
}
/// Spawn a background task that syncs model metadata immediately
/// and then every 10 minutes. Returns the `JoinHandle`.
///
/// Failures are logged but do not stop the task — it keeps retrying.
pub fn start_sync_task(self) -> JoinHandle<()> {
let db = self.db.clone();
let ai_api_key = self.config.ai_api_key().ok();
let ai_base_url = self.config.ai_basic_url().ok();
tokio::spawn(async move {
// Run once immediately on startup before taking traffic.
Self::sync_once(&db, ai_api_key.clone(), ai_base_url.clone()).await;
let mut tick = interval(Duration::from_secs(60 * 10));
loop {
tick.tick().await;
Self::sync_once(&db, ai_api_key.clone(), ai_base_url.clone()).await;
}
})
}
/// Perform a single sync pass. Errors are logged and silently swallowed
/// so the periodic task never stops.
async fn sync_once(db: &AppDatabase, ai_api_key: Option<String>, ai_base_url: Option<String>) {
let (http_client, base_url, api_key) =
match build_ai_client_from_parts(ai_api_key, ai_base_url) {
Ok(c) => c,
Err(msg) => {
tracing::warn!(error = %msg, "Model sync: AI client config error");
return;
}
};
let upstream_models = match list_upstream_models(&http_client, &base_url, &api_key).await {
Ok(models) => models,
Err(e) => {
tracing::warn!(error = ?e, "Model sync: failed to list upstream models");
return;
}
};
tracing::info!(
model_count = upstream_models.len(),
"Model sync: {} models from upstream",
upstream_models.len()
);
let result = sync_models_from_upstream(db, upstream_models).await;
tracing::info!(
models_created = result.models_created,
models_updated = result.models_updated,
versions_created = result.versions_created,
pricing_created = result.pricing_created,
capabilities_created = result.capabilities_created,
profiles_created = result.profiles_created,
"Model sync complete"
);
}
}