gitdataai/libs/agent/embed/client.rs

use async_openai::Client;
use async_openai::types::embeddings::CreateEmbeddingRequestArgs;
use serde::{Deserialize, Serialize};

use crate::embed::qdrant::QdrantClient;

pub struct EmbedClient {
    openai: Client<async_openai::config::OpenAIConfig>,
    qdrant: QdrantClient,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct EmbedVector {
    pub id: String,
    pub vector: Vec<f32>,
    pub payload: EmbedPayload,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct EmbedPayload {
    pub entity_type: String,
    pub entity_id: String,
    pub text: String,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub extra: Option<serde_json::Value>,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct SearchResult {
    pub id: String,
    pub score: f32,
    pub payload: EmbedPayload,
}

impl EmbedClient {
    pub fn new(openai: Client<async_openai::config::OpenAIConfig>, qdrant: QdrantClient) -> Self {
        Self { openai, qdrant }
    }

    pub async fn embed_text(&self, text: &str, model: &str) -> crate::Result<Vec<f32>> {
        let request = CreateEmbeddingRequestArgs::default()
            .model(model)
            .input(text)
            .build()
            .map_err(|e| crate::AgentError::OpenAi(e.to_string()))?;

        let response = self
            .openai
            .embeddings()
            .create(request)
            .await
            .map_err(|e| crate::AgentError::OpenAi(e.to_string()))?;

        response
            .data
            .first()
            .map(|d| d.embedding.clone())
            .ok_or_else(|| crate::AgentError::OpenAi("no embedding returned".into()))
    }

    pub async fn embed_batch(&self, texts: &[String], model: &str) -> crate::Result<Vec<Vec<f32>>> {
        let request = CreateEmbeddingRequestArgs::default()
            .model(model)
            .input(texts.to_vec())
            .build()
            .map_err(|e| crate::AgentError::OpenAi(e.to_string()))?;

        let response = self
            .openai
            .embeddings()
            .create(request)
            .await
            .map_err(|e| crate::AgentError::OpenAi(e.to_string()))?;

        let mut embeddings = vec![Vec::new(); texts.len()];
        for data in response.data {
            if (data.index as usize) < embeddings.len() {
                embeddings[data.index as usize] = data.embedding;
            }
        }
        Ok(embeddings)
    }

    pub async fn upsert(&self, points: Vec<EmbedVector>) -> crate::Result<()> {
        self.qdrant.upsert_points(points).await
    }

    pub async fn search(
        &self,
        query: &str,
        entity_type: &str,
        model: &str,
        limit: usize,
    ) -> crate::Result<Vec<SearchResult>> {
        let vector = self.embed_text(query, model).await?;
        self.qdrant.search(&vector, entity_type, limit).await
    }

    pub async fn search_with_filter(
        &self,
        query: &str,
        entity_type: &str,
        model: &str,
        limit: usize,
        filter: qdrant_client::qdrant::Filter,
    ) -> crate::Result<Vec<SearchResult>> {
        let vector = self.embed_text(query, model).await?;
        self.qdrant
            .search_with_filter(&vector, entity_type, limit, filter)
            .await
    }

    pub async fn delete_by_entity_id(
        &self,
        entity_type: &str,
        entity_id: &str,
    ) -> crate::Result<()> {
        self.qdrant.delete_by_filter(entity_type, entity_id).await
    }

    pub async fn ensure_collection(&self, entity_type: &str, dimensions: u64) -> crate::Result<()> {
        self.qdrant.ensure_collection(entity_type, dimensions).await
    }

    pub async fn ensure_memory_collection(&self, dimensions: u64) -> crate::Result<()> {
        self.qdrant.ensure_memory_collection(dimensions).await
    }

    pub async fn ensure_skill_collection(&self, dimensions: u64) -> crate::Result<()> {
        self.qdrant.ensure_skill_collection(dimensions).await
    }

    /// Embed and store a conversation memory (message) in Qdrant.
    pub async fn embed_memory(
        &self,
        id: &str,
        text: &str,
        room_id: &str,
        user_id: Option<&str>,
        model: &str,
    ) -> crate::Result<()> {
        let vector = self.embed_text(text, model).await?;
        let point = EmbedVector {
            id: id.to_string(),
            vector,
            payload: EmbedPayload {
                entity_type: "memory".to_string(),
                entity_id: room_id.to_string(),
                text: text.to_string(),
                extra: serde_json::json!({ "user_id": user_id }).into(),
            },
        };
        self.qdrant.upsert_points(vec![point]).await
    }

    /// Search memory embeddings by semantic similarity within a room.
    pub async fn search_memories(
        &self,
        query: &str,
        model: &str,
        room_id: &str,
        limit: usize,
    ) -> crate::Result<Vec<SearchResult>> {
        let vector = self.embed_text(query, model).await?;
        let mut results = self.qdrant.search_memory(&vector, limit + 1).await?;
        // Filter to the specific room
        results.retain(|r| r.payload.entity_id == room_id);
        results.truncate(limit);
        Ok(results)
    }

    /// Embed and store a skill in Qdrant.
    pub async fn embed_skill(
        &self,
        id: &str,
        name: &str,
        description: &str,
        content: &str,
        project_uuid: &str,
        model: &str,
    ) -> crate::Result<()> {
        let text = format!("{}: {} {}", name, description, content);
        let vector = self.embed_text(&text, model).await?;
        let point = EmbedVector {
            id: id.to_string(),
            vector,
            payload: EmbedPayload {
                entity_type: "skill".to_string(),
                entity_id: project_uuid.to_string(),
                text,
                extra: serde_json::json!({ "name": name, "description": description }).into(),
            },
        };
        self.qdrant.upsert_points(vec![point]).await
    }

    /// Search skill embeddings by semantic similarity within a project.
    pub async fn search_skills(
        &self,
        query: &str,
        model: &str,
        project_uuid: &str,
        limit: usize,
    ) -> crate::Result<Vec<SearchResult>> {
        let vector = self.embed_text(query, model).await?;
        let mut results = self.qdrant.search_skill(&vector, limit + 1).await?;
        results.retain(|r| r.payload.entity_id == project_uuid);
        results.truncate(limit);
        Ok(results)
    }
}