Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors
Rust
fix(embed): bound the embedding input on the Gemini and Voyage embedders
Rudy Mizrahi Celeklia8d13041916e
diff --git a/crates/ai-memory-llm/src/embedding.rs b/crates/ai-memory-llm/src/embedding.rsindex ae985acb..134f5b45 100644--- a/crates/ai-memory-llm/src/embedding.rs+++ b/crates/ai-memory-llm/src/embedding.rs@@ -677,8 +677,9 @@ impl Embedder for VoyageEmbedder {async fn embed(&self, text: &str) -> LlmResult<Vec<f32>> {let url = normalize_openai_base(&self.base_url, "embeddings");+ let input = truncate_for_embedding(text, OPENAI_EMBED_MAX_TOKENS);let req = VoyageRequest {- input: [text],+ input: [input.as_str()],model: &self.model,};let mut attempt = 0u32;diff --git a/crates/ai-memory-llm/src/google.rs b/crates/ai-memory-llm/src/google.rsindex 0bdb3e5a..87e1fdc2 100644--- a/crates/ai-memory-llm/src/google.rs+++ b/crates/ai-memory-llm/src/google.rs@@ -9,9 +9,10 @@ use secrecy::{ExposeSecret, SecretString};use serde::{Deserialize, Serialize};use tracing::debug;-use crate::embedding::{Embedder, normalise, retry_delay};+use crate::embedding::{Embedder, OPENAI_EMBED_MAX_TOKENS, normalise, retry_delay};use crate::error::{LlmError, LlmResult};use crate::response::{provider_error_body, response_json_limited};+use crate::text::truncate_for_embedding;/// Default Gemini API host.pub const DEFAULT_BASE_URL: &str = "https://generativelanguage.googleapis.com";@@ -74,6 +75,10 @@ impl GoogleEmbedder {} else {text.to_string()};+ // Gemini rejects an over-long `embedContent` input with 400+ // INVALID_ARGUMENT, and `embed_page_version` only logs that failure, so+ // an unbounded body would leave the page with no vector at all.+ let prepared = truncate_for_embedding(&prepared, OPENAI_EMBED_MAX_TOKENS);let url = embed_url(&self.base_url, &self.model);let body = GeminiEmbedRequest {@@ -262,6 +267,58 @@ mod tests {assert!(format_query_v2("find auth").contains("query: find auth"));}+ #[derive(Clone)]+ struct AssertBoundedInput;++ impl Respond for AssertBoundedInput {+ fn respond(&self, req: &Request) -> ResponseTemplate {+ let body: serde_json::Value = match serde_json::from_slice(&req.body) {+ Ok(value) => value,+ Err(err) => {+ return ResponseTemplate::new(500).set_body_string(format!("bad body: {err}"));+ }+ };+ let text = body["content"]["parts"][0]["text"]+ .as_str()+ .unwrap_or_default();+ if text.len() > 8_000 {+ return ResponseTemplate::new(500).set_body_string(format!(+ "unbounded embedContent input: {} bytes",+ text.len()+ ));+ }+ if !text.ends_with('\u{2026}') {+ return ResponseTemplate::new(500)+ .set_body_string("expected the truncation ellipsis");+ }+ ResponseTemplate::new(200).set_body_json(serde_json::json!({+ "embedding": { "values": [1.0, 0.0, 0.0] }+ }))+ }+ }++ #[tokio::test]+ async fn embed_content_bounds_a_long_input() {+ let server = MockServer::start().await;+ Mock::given(method("POST"))+ .and(path("/v1beta/models/gemini-embedding-001:embedContent"))+ .respond_with(AssertBoundedInput)+ .mount(&server)+ .await;++ let embedder =+ GoogleEmbedder::new(SecretString::from("test-key"), "gemini-embedding-001", 3)+ .expect("google embedder builds")+ .with_base_url(server.uri());++ let embedding = embedder+ .embed_document(&"word ".repeat(8_000))+ .await+ .expect("a long page is bounded before the request");++ assert_eq!(embedding, vec![1.0, 0.0, 0.0]);+ }+#[tokio::test]async fn embed_content_uses_api_key_header_not_bearer_auth() {let server = MockServer::start().await;