Skip to content

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

fix(tinyhumans): guard managed embeddings at the transport boundary

Steven Enamakel3c0eeed5f2f6
diff --git a/crates/openhuman-tinyhumans/src/inference/managed_embedding.rs b/crates/openhuman-tinyhumans/src/inference/managed_embedding.rs
index 555f28a88a..afad4601e9 100644
--- a/crates/openhuman-tinyhumans/src/inference/managed_embedding.rs
+++ b/crates/openhuman-tinyhumans/src/inference/managed_embedding.rs
@@ -7,7 +7,7 @@ use std::sync::Arc;
use openhuman_embed::__host::egress::{emit_external_transfer, enforce_egress, EgressDescriptor};
use openhuman_embed::__host::inference_registry::EmbeddingRequest;
use openhuman_embed::__host::security::approval::ApprovalChatContext;
-use tinyinference_embeddings::{BearerResolver, CloudEmbeddingModel, EmbeddingModel};
+use tinyinference_embeddings::{BearerResolver, CloudEmbeddingModel, EmbeddingEgressGuard, EmbeddingModel};
use super::managed_session::resolve_session_bearer;
use crate::backend::url::effective_api_url;
@@ -21,8 +21,11 @@ use crate::client::endpoint::is_managed_endpoint_for_api_key;
/// chat of the turn that made the call. One built without a chat (background
/// sync, the embedding RPC) still discloses, unrouted.
pub struct ManagedEmbeddingModel {
- inner: Arc<CloudEmbeddingModel>,
+ inner: CloudEmbeddingModel,
+ base_url: String,
model: String,
+ dimensions: usize,
+ bearer: BearerResolver,
chat: Option<ApprovalChatContext>,
}
@@ -82,14 +85,13 @@ impl ManagedEmbeddingModel {
.map_err(|error| tinyinference_embeddings::Error::Embedding(error.to_string()))
});
let model = model.to_string();
+ let base_url = base_url;
Self {
- inner: Arc::new(CloudEmbeddingModel::new(
- base_url,
- model.clone(),
- dimensions,
- bearer,
- )),
+ inner: cloud_model(&base_url, &model, dimensions, Arc::clone(&bearer), chat.as_ref()),
+ base_url,
model,
+ dimensions,
+ bearer,
chat,
}
}
@@ -101,19 +103,40 @@ impl ManagedEmbeddingModel {
#[must_use]
pub fn routed_to(&self, chat: Option<ApprovalChatContext>) -> Self {
Self {
- inner: Arc::clone(&self.inner),
+ inner: cloud_model(
+ &self.base_url,
+ &self.model,
+ self.dimensions,
+ Arc::clone(&self.bearer),
+ chat.as_ref(),
+ ),
+ base_url: self.base_url.clone(),
model: self.model.clone(),
+ dimensions: self.dimensions,
+ bearer: Arc::clone(&self.bearer),
chat,
}
}
+}
- fn guard(&self) -> tinyinference_embeddings::Result<()> {
- let egress = EgressDescriptor::embedding("cloud", &self.model);
+fn cloud_model(
+ base_url: &str,
+ model: &str,
+ dimensions: usize,
+ bearer: BearerResolver,
+ chat: Option<&ApprovalChatContext>,
+) -> CloudEmbeddingModel {
+ let guarded_model = model.to_owned();
+ let guarded_chat = chat.cloned();
+ let egress_guard: EmbeddingEgressGuard = Arc::new(move |_, _| {
+ let egress = EgressDescriptor::embedding("cloud", &guarded_model);
enforce_egress(&egress)
.map_err(|error| tinyinference_embeddings::Error::Embedding(error.to_string()))?;
- emit_external_transfer(self.chat.as_ref(), egress);
+ emit_external_transfer(guarded_chat.as_ref(), egress);
Ok(())
- }
+ });
+ CloudEmbeddingModel::new(base_url.to_owned(), model.to_owned(), dimensions, bearer)
+ .with_egress_guard(egress_guard)
}
#[async_trait::async_trait]
@@ -135,11 +158,8 @@ impl EmbeddingModel for ManagedEmbeddingModel {
}
async fn embed(&self, texts: &[String]) -> tinyinference_embeddings::Result<Vec<Vec<f32>>> {
- // Only a batch that will actually leave is guarded: an empty one answers
- // locally, and a blank input is refused by the transport.
- if !texts.is_empty() && texts.iter().all(|text| !text.trim().is_empty()) {
- self.guard()?;
- }
+ // The owning transport invokes the host guard after input validation
+ // and immediately before resolving credentials or sending a request.
self.inner.embed(texts).await
}
}
diff --git a/crates/openhuman-tinyhumans/src/inference/managed_model.rs b/crates/openhuman-tinyhumans/src/inference/managed_model.rs
index f905f8a5c7..5180c37e8c 100644
--- a/crates/openhuman-tinyhumans/src/inference/managed_model.rs
+++ b/crates/openhuman-tinyhumans/src/inference/managed_model.rs
@@ -73,7 +73,7 @@ pub(super) fn build_chat_model(
let mut managed = TinyHumansModel::new(
endpoint,
model,
- model,
+ super::tiers::MODEL_MANAGED_DEFAULT,
move || {
resolver_options
.resolve_bearer(&resolver_endpoint)
@@ -93,6 +93,10 @@ pub(super) fn build_chat_model(
Ok(managed)
}
+#[cfg(test)]
+#[path = "managed_model_tests.rs"]
+mod tests;
+
/// Connect timeout for managed inference; matches the previous host adapter.
const MANAGED_INFERENCE_CONNECT_TIMEOUT: Duration = Duration::from_secs(30);
/// Keep pooled connections below the backend ALB's idle timeout.
diff --git a/crates/openhuman-tinyhumans/src/inference/managed_model_tests.rs b/crates/openhuman-tinyhumans/src/inference/managed_model_tests.rs
index 8f3e7b0496..c0837600dd 100644
--- a/crates/openhuman-tinyhumans/src/inference/managed_model_tests.rs
+++ b/crates/openhuman-tinyhumans/src/inference/managed_model_tests.rs
@@ -18,3 +18,34 @@ fn managed_model_preserves_the_openhuman_response_cache_identity() {
Some("openhuman:https://api.example.test/openai/v1:reasoning-v1")
);
}
+
+#[test]
+fn managed_model_trims_a_configured_model_and_uses_the_product_default_when_blank() {
+ let configured = build_chat_model(
+ Some("https://api.example.test"),
+ &ManagedOptions::default(),
+ " reasoning-v1 ",
+ true,
+ None,
+ false,
+ )
+ .expect("configured model");
+ assert_eq!(
+ configured.cache_identity().as_deref(),
+ Some("openhuman:https://api.example.test/openai/v1:reasoning-v1")
+ );
+
+ let blank = build_chat_model(
+ Some("https://api.example.test"),
+ &ManagedOptions::default(),
+ " ",
+ true,
+ None,
+ false,
+ )
+ .expect("blank model falls back to the product default");
+ assert_eq!(
+ blank.cache_identity().as_deref(),
+ Some("openhuman:https://api.example.test/openai/v1:reasoning-v1")
+ );
+}