feat(core): implement full memory pipeline (#11)

crates/mem-llm/Cargo.toml

@@ -14,3 +14,5 @@ thiserror = { workspace = true }
 reqwest = { workspace = true }
 tracing = { workspace = true }
 chrono = { workspace = true }
+pgvector = { workspace = true }
+uuid = { workspace = true }

crates/mem-llm/src/chat.rs

@@ -144,10 +144,13 @@ impl ChatClient {
         let mut last_error: Option<anyhow::Error> = None;
 
         for attempt in 0..self.max_retries {
-            let response = self
-                .http
-                .post(&url)
-                .header("apikey", &self.api_key)
+            let mut req = self.http.post(&url);
+            // Only add apikey header if it's not empty (for backward compatibility)
+            if !self.api_key.is_empty() && !self.api_key.starts_with("http") {
+                req = req.header("apikey", &self.api_key);
+            }
+            
+            let response = req
                 .header("Content-Type", "application/json")
                 .body(body.clone())
                 .timeout(self.timeout)

crates/mem-llm/src/embeddings.rs

@@ -0,0 +1,64 @@
+use anyhow::{anyhow, Result};
+use pgvector::Vector;
+use reqwest::Client;
+use serde::{Deserialize, Serialize};
+use std::env;
+
+/// Embeddings client for Ollama
+#[derive(Clone)]
+pub struct EmbeddingsClient {
+    base_url: String,
+    model: String,
+    #[allow(dead_code)]
+    http: Client,
+}
+
+#[derive(Debug, Serialize)]
+struct EmbeddingRequest {
+    model: String,
+    input: Vec<String>,
+}
+
+#[derive(Debug, Deserialize)]
+struct EmbeddingResponse {
+    embeddings: Vec<Vec<f32>>,
+    model: String,
+}
+
+impl EmbeddingsClient {
+    /// Create from environment
+    /// Uses api.riotpiao.com gateway (nomic-ai/nomic-embed-text-v2-moe model)
+    pub fn from_env() -> Result<Self> {
+        let base_url = env::var("LLM_API_BASE").unwrap_or_else(|_| "https://api.riotpiao.com".to_string());
+        let model = "nomic-ai/nomic-embed-text-v2-moe".to_string();
+
+        Ok(Self {
+            base_url,
+            model,
+            http: Client::new(),
+        })
+    }
+
+    /// Embed a single text string
+    pub async fn embed(&self, text: &str) -> Result<Vector> {
+        let embeddings = self.embed_batch(&[text.to_string()]).await?;
+        Ok(embeddings.into_iter().next().ok_or_else(|| anyhow::anyhow!("empty embedding response"))?)
+    }
+
+    /// Embed multiple texts in a batch using api.riotpiao.com gateway
+    pub async fn embed_batch(&self, texts: &[String]) -> Result<Vec<Vector>> {
+        let req = EmbeddingRequest {
+            model: self.model.clone(),
+            input: texts.to_vec(),
+        };
+
+        let url = format!("{}/v1/embeddings", self.base_url);
+        let resp: EmbeddingResponse = self.http.post(&url).json(&req).send().await?.json().await?;
+
+        Ok(resp
+            .embeddings
+            .into_iter()
+            .map(Vector::from)
+            .collect())
+    }
+}

crates/mem-llm/src/lib.rs

@@ -1,5 +1,7 @@
 pub mod chat;
 pub mod rerank;
+pub mod embeddings;
 
 pub use chat::{ChatClient, Completion, Usage};
 pub use rerank::RerankClient;
+pub use embeddings::EmbeddingsClient;

crates/mem-llm/src/rerank.rs

@@ -1,33 +1,51 @@
 use anyhow::Result;
 use reqwest::Client;
-use serde_json::json;
+use serde::{Deserialize, Serialize};
+use std::time::Duration;
 
-/// Rerank response item (bare array, not OpenAI envelope).
-#[derive(serde::Deserialize, Debug)]
+/// Rerank score result
+#[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct RerankScore {
     pub index: usize,
     pub score: f32,
 }
 
-/// Rerank client (BAAI/bge-reranker-base via TEI).
+/// Rerank response from gateway
+#[derive(Deserialize)]
+struct RerankResponse {
+    results: Vec<RerankScore>,
+}
+
+/// Rerank client using api.riotpiao.com gateway (BAAI/bge-reranker-base model)
 pub struct RerankClient {
     base_url: String,
-    api_key: String,
     model: String,
     timeout_secs: u64,
 }
 
 impl RerankClient {
-    /// Create rerank client.
-    pub fn new(base_url: &str, api_key: &str, model: &str) -> Result<Self> {
+    /// Create rerank client pointing to gateway
+    pub fn new(base_url: &str, _api_key: &str, model: &str) -> Result<Self> {
         Ok(Self {
             base_url: base_url.to_string(),
-            api_key: api_key.to_string(),
             model: model.to_string(),
             timeout_secs: 300,
         })
     }
 
+    /// Create from environment (uses api.riotpiao.com)
+    pub fn from_env() -> Result<Self> {
+        let base_url = std::env::var("LLM_API_BASE")
+            .unwrap_or_else(|_| "https://api.riotpiao.com".to_string());
+        let model = "BAAI/bge-reranker-base".to_string();
+        
+        Ok(Self {
+            base_url,
+            model,
+            timeout_secs: 300,
+        })
+    }
+
     /// Rerank query against texts, return scored items in score order.
     /// Returns Vec<(index, score)> mapping back to input positions.
     pub async fn rerank(&self, query: &str, texts: &[&str]) -> Result<Vec<(usize, f32)>> {
@@ -36,39 +54,41 @@ impl RerankClient {
             return Ok(vec![]);
         }
 
-        let url = format!("{}/rerank", self.base_url);
+        let url = format!("{}/v1/rerank", self.base_url);
 
         let client = Client::builder()
-            .timeout(std::time::Duration::from_secs(self.timeout_secs))
+            .timeout(Duration::from_secs(self.timeout_secs))
             .build()?;
 
-        let payload = json!({
+        let payload = serde_json::json!({
+            "model": self.model,
             "query": query,
             "texts": texts,
+            "top_k": texts.len(),
         });
 
         let response = client
             .post(&url)
-            .header("apikey", &self.api_key)
             .header("Content-Type", "application/json")
             .json(&payload)
             .send()
             .await?;
 
         if !response.status().is_success() {
-            return Err(anyhow::anyhow!("Rerank failed: {}", response.status()));
+            let error_text = response.text().await.unwrap_or_default();
+            return Err(anyhow::anyhow!("Rerank failed: {}", error_text));
         }
 
-        // Parse bare array (not OpenAI envelope)
-        let scores: Vec<RerankScore> = response.json().await?;
+        // Parse gateway response (OpenAI format with results field)
+        let resp: RerankResponse = response.json().await?;
 
-        // Map back to input positions and scores
-        let mut results: Vec<(usize, f32)> = scores
+        // Map to (index, score) and sort by score descending
+        let mut results: Vec<(usize, f32)> = resp
+            .results
             .into_iter()
             .map(|s| (s.index, s.score))
             .collect();
 
-        // Sort by score descending (highest first)
         results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
 
         Ok(results)