Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

refactor(embedding_host): move managed embedder fallback into tinyinference

Steven Enamakel4f25e093b64f
diff --git a/crates/openhuman-core/src/inference/embedding_host/cloud_adapter.rs b/crates/openhuman-core/src/inference/embedding_host/cloud_adapter.rs
deleted file mode 100644
index eef1c7d0ba..0000000000
--- a/crates/openhuman-core/src/inference/embedding_host/cloud_adapter.rs
+++ /dev/null
@@ -1,56 +0,0 @@
-//! Unavailable placeholder for the managed cloud embedder. The embedder itself is a registered provider
-//! ([`crate::inference::registry::register_embedding`]); with none registered
-//! the managed embedder is a placeholder that answers `BACKEND_UNAVAILABLE:`.
-
-use tinyinference_embeddings::{EmbeddingModel, DEFAULT_CLOUD_DIMENSIONS, DEFAULT_CLOUD_MODEL};
-
-pub const DEFAULT_CLOUD_EMBEDDING_MODEL: &str = DEFAULT_CLOUD_MODEL;
-pub const DEFAULT_CLOUD_EMBEDDING_DIMENSIONS: usize = DEFAULT_CLOUD_DIMENSIONS;
-
-/// The managed embedder when no provider is registered for it: keeps the
-/// configured identity (name, model, dimensions) so callers that only describe
-/// the embedder still work, and refuses to embed with the `BACKEND_UNAVAILABLE:`
-/// sentinel.
-pub struct UnavailableEmbeddingModel {
- model: String,
- dimensions: usize,
-}
-
-impl UnavailableEmbeddingModel {
- pub fn new(model: &str, dimensions: usize) -> Self {
- Self {
- model: model.to_string(),
- dimensions,
- }
- }
-}
-
-#[async_trait::async_trait]
-impl EmbeddingModel for UnavailableEmbeddingModel {
- fn name(&self) -> &str {
- "cloud"
- }
-
- fn model_id(&self) -> &str {
- &self.model
- }
-
- fn dimensions(&self) -> usize {
- self.dimensions
- }
-
- fn signature(&self) -> String {
- format!("unavailable:{}:{}", self.model, self.dimensions)
- }
-
- async fn embed(&self, _texts: &[String]) -> tinyinference_embeddings::Result<Vec<Vec<f32>>> {
- Err(tinyinference_embeddings::Error::Validation(format!(
- "{} managed embeddings need a hosted provider; none is registered",
- crate::core::observability::BACKEND_UNAVAILABLE_PREFIX
- )))
- }
-}
-
-#[cfg(test)]
-#[path = "cloud_adapter_tests.rs"]
-mod tests;
diff --git a/crates/openhuman-core/src/inference/embedding_host/cloud_adapter_tests.rs b/crates/openhuman-core/src/inference/embedding_host/cloud_adapter_tests.rs
deleted file mode 100644
index b7f56dfc8c..0000000000
--- a/crates/openhuman-core/src/inference/embedding_host/cloud_adapter_tests.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-use super::*;
-
-/// With no embedding provider registered, the managed embedder keeps its
-/// identity and answers with the `BACKEND_UNAVAILABLE:` sentinel on use.
-#[tokio::test]
-async fn unregistered_managed_embedder_answers_backend_unavailable() {
- let provider = UnavailableEmbeddingModel::new(
- DEFAULT_CLOUD_EMBEDDING_MODEL,
- DEFAULT_CLOUD_EMBEDDING_DIMENSIONS,
- );
- assert_eq!(provider.name(), "cloud");
- assert_eq!(provider.dimensions(), DEFAULT_CLOUD_EMBEDDING_DIMENSIONS);
- let err = provider.embed(&["hello".to_string()]).await.unwrap_err();
- assert!(
- err.to_string()
- .contains(crate::core::observability::BACKEND_UNAVAILABLE_PREFIX),
- "unexpected error: {err}"
- );
-}
diff --git a/crates/openhuman-core/src/inference/embedding_host/factory.rs b/crates/openhuman-core/src/inference/embedding_host/factory.rs
index 49a5c85f64..08706662cb 100644
--- a/crates/openhuman-core/src/inference/embedding_host/factory.rs
+++ b/crates/openhuman-core/src/inference/embedding_host/factory.rs
@@ -77,22 +77,6 @@ pub fn create_embedding_provider_with_credentials(
.unwrap_or_else(|| provider.to_owned());
let api_key = api_key.to_owned();
let credentials: CredentialResolver = Arc::new(move |_| api_key.clone());
- let managed: ManagedModelFactory = Arc::new(|model, dimensions| {
- if let Some(registered) = crate::inference::registry::embedding(PROVIDER_OPENHUMAN) {
- log::debug!("[embeddings::factory] managed model served by a registered factory");
- let (state_dir, secrets_encrypt) = credential_scope(None);
- return registered(&EmbeddingRequest {
- config: None,
- model,
- dimensions,
- chat: None,
- state_dir,
- secrets_encrypt,
- });
- }
- log::debug!("[embeddings::factory] no managed embedding provider registered");
- Ok(Box::new(UnavailableEmbeddingModel::new(model, dimensions)))
- });
let model = create_configured_embedding_model(
&settings(
&provider,
@@ -101,7 +85,7 @@ pub fn create_embedding_provider_with_credentials(
tinyinference_local::ollama::ollama_base_url(),
),
&credentials,
- &managed,
+ &managed_factory(None, None),
)?;
Ok(model)
}
@@ -153,8 +137,11 @@ pub fn create_embedding_provider_for_chat(
);
let api_key = api_key.to_owned();
let credentials: CredentialResolver = Arc::new(move |_| api_key.clone());
- let model =
- create_configured_embedding_model(&selected, &credentials, &managed_factory(config, chat))?;
+ let model = create_configured_embedding_model(
+ &selected,
+ &credentials,
+ &managed_factory(Some(config.clone()), chat),
+ )?;
Ok(model)
}
@@ -181,7 +168,7 @@ pub fn default_embedding_provider_for_chat(
let model = create_default_embedding_model(
&selected,
&credential_resolver(config),
- &managed_factory(config, chat),
+ &managed_factory(Some(config.clone()), chat),
)
.unwrap_or_else(|error| {
log::warn!("[embeddings::factory] managed fallback failed to build: {error}");
diff --git a/crates/openhuman-core/src/inference/embedding_host/factory_tests.rs b/crates/openhuman-core/src/inference/embedding_host/factory_tests.rs
index 7e80adeb02..e527affae9 100644
--- a/crates/openhuman-core/src/inference/embedding_host/factory_tests.rs
+++ b/crates/openhuman-core/src/inference/embedding_host/factory_tests.rs
@@ -1,6 +1,7 @@
use super::*;
-use crate::inference::embedding_host::cloud::{
- DEFAULT_CLOUD_EMBEDDING_DIMENSIONS, DEFAULT_CLOUD_EMBEDDING_MODEL,
+use tinyinference_embeddings::{
+ DEFAULT_CLOUD_DIMENSIONS as DEFAULT_CLOUD_EMBEDDING_DIMENSIONS,
+ DEFAULT_CLOUD_MODEL as DEFAULT_CLOUD_EMBEDDING_MODEL,
};
use crate::security::credentials::{AuthService, APP_SESSION_PROVIDER};
use std::collections::HashMap;
diff --git a/crates/openhuman-core/src/inference/embedding_host/mod.rs b/crates/openhuman-core/src/inference/embedding_host/mod.rs
index 10eb9a288a..fe45bac5b5 100644
--- a/crates/openhuman-core/src/inference/embedding_host/mod.rs
+++ b/crates/openhuman-core/src/inference/embedding_host/mod.rs
@@ -13,13 +13,10 @@
//! - **Custom**: Any OpenAI-compatible endpoint.
//! - **Noop**: A fallback provider for keyword-only search.
-#[path = "cloud_adapter.rs"]
-pub mod cloud;
mod factory;
mod rpc;
mod schemas;
-pub use cloud::UnavailableEmbeddingModel;
pub use factory::{
create_embedding_provider_for_chat, create_embedding_provider_with_config,
create_embedding_provider_with_credentials, default_embedding_provider_for_chat,
diff --git a/crates/openhuman-core/src/inference/embedding_host/mod_tests.rs b/crates/openhuman-core/src/inference/embedding_host/mod_tests.rs
index 675558199d..529187b1b8 100644
--- a/crates/openhuman-core/src/inference/embedding_host/mod_tests.rs
+++ b/crates/openhuman-core/src/inference/embedding_host/mod_tests.rs
@@ -4,7 +4,10 @@
//! `#[path = "mod_tests.rs"] mod tests;` so `super::*` still resolves to the
//! `embeddings` module.
-use super::cloud::{DEFAULT_CLOUD_EMBEDDING_DIMENSIONS, DEFAULT_CLOUD_EMBEDDING_MODEL};
+use tinyinference_embeddings::{
+ DEFAULT_CLOUD_DIMENSIONS as DEFAULT_CLOUD_EMBEDDING_DIMENSIONS,
+ DEFAULT_CLOUD_MODEL as DEFAULT_CLOUD_EMBEDDING_MODEL,
+};
use super::*;
// ── create_embedding_provider_with_credentials ───────────
diff --git a/crates/openhuman-core/src/inference/embedding_host/rpc/settings.rs b/crates/openhuman-core/src/inference/embedding_host/rpc/settings.rs
index 48a69af08b..f50d8d45b4 100644
--- a/crates/openhuman-core/src/inference/embedding_host/rpc/settings.rs
+++ b/crates/openhuman-core/src/inference/embedding_host/rpc/settings.rs
@@ -9,15 +9,11 @@ use tinyinference_providers::route::{custom_provider_string, split_custom_provid
use super::{resolve_api_key, LOG_PREFIX};
use tinyinference_embeddings::catalog;
-use tinyinference_embeddings::probe::{
- classify_embed_probe, final_probe_dims, probe_custom_embeddings, EmbedProbe,
- EmbeddingProbeRejection,
-};
-use tinyinference_embeddings::served_models::{
- check_requested_model_served, fetch_served_model_ids, ModelNotServed,
+use tinyinference_embeddings::verify::{
+ effective_embedder_slug, verify_custom_endpoint, VerifyRejection,
};
-fn probe_rejection_outcome(reject: EmbeddingProbeRejection) -> Outcome<serde_json::Value> {
+fn probe_rejection_outcome(reject: tinyinference_embeddings::probe::EmbeddingProbeRejection) -> Outcome<serde_json::Value> {
let mut body = serde_json::json!({ "error": reject.error, "message": reject.message });
if let Some(detail) = reject.detail {
body["detail"] = serde_json::Value::String(detail);
@@ -25,7 +21,7 @@ fn probe_rejection_outcome(reject: EmbeddingProbeRejection) -> Outcome<serde_jso
Outcome::new(body, vec![reject.summary.to_string()])
}
-fn model_rejection_outcome(reject: ModelNotServed) -> Outcome<serde_json::Value> {
+fn model_rejection_outcome(reject: tinyinference_embeddings::served_models::ModelNotServed) -> Outcome<serde_json::Value> {
let mut body = serde_json::json!({
"error": reject.error,
"message": reject.message,
@@ -38,49 +34,6 @@ fn model_rejection_outcome(reject: ModelNotServed) -> Outcome<serde_json::Value>
Outcome::new(body, vec![reject.summary.to_string()])
}
-/// Slug naming the embedder ingestion will actually use, resolved host-side
-/// from the `Config` fields the resolution ladder reads.
-///
-/// Resolution order:
-/// 1. Deliberate opt-out — `embeddings_provider` trimmed equals `"none"` → `"none"`.
-/// 2. Local Ollama via unified workload setting — `workload_local_model("embeddings")`
-/// is `Some` → `"ollama"`.
-/// 3. User OpenAI-compatible endpoint — `memory.embedding_provider` is
-/// `"openai"`, `"custom"`, or starts with `"custom:"` → `"custom"`.
-/// 4. Managed cloud session — `auth-profiles.json` exists next to the config
-/// file → `"cloud"`.
-/// 5. Nothing usable → `"unconfigured"`.
-fn effective_embedder_slug_from_config(config: &Config) -> &'static str {
- // 1. Deliberate opt-out.
- if config
- .embeddings_provider
- .as_deref()
- .map(str::trim)
- .is_some_and(|s| s == "none")
- {
- return "none";
- }
- // 2. Local Ollama via unified workload setting.
- if config.workload_local_model("embeddings").is_some() {
- return "ollama";
- }
- // 3. User OpenAI-compatible endpoint.
- let picker = config.memory.embedding_provider.trim();
- if picker == "openai" || split_custom_provider(picker).0 == "custom" {
- return "custom";
- }
- // 4. Managed cloud session.
- let session_exists = config
- .config_path
- .parent()
- .map(|dir| dir.join("auth-profiles.json").exists())
- .unwrap_or(false);
- if session_exists {
- return "cloud";
- }
- "unconfigured"
-}
-
fn active_custom_profile(config: &Config) -> Option<crate::config::schema::CustomEmbeddingsConfig> {
let (_, endpoint) = split_custom_provider(&config.memory.embedding_provider);
let endpoint = endpoint?;
@@ -162,7 +115,22 @@ pub async fn get_settings(config: &Config) -> Result<Outcome<serde_json::Value>,
// routes to local Ollama without rewriting it. Additive field — callers that only need the picker
// value are unaffected; callers asking "does this bill the managed budget?"
// must read this one (#5402).
- let effective_provider = effective_embedder_slug_from_config(config);
+ let local_model = config.workload_local_model("embeddings");
+ let session_exists = config
+ .config_path
+ .parent()
+ .map(|dir| dir.join("auth-profiles.json").exists())
+ .unwrap_or(false);
+ let effective_provider = effective_embedder_slug(
+ config
+ .embeddings_provider
+ .as_deref()
+ .map(str::trim)
+ .is_some_and(|provider| provider == "none"),
+ local_model.as_deref(),
+ Some(&config.memory.embedding_provider),
+ session_exists,
+ );
let payload = serde_json::json!({
"provider": provider,