feat: Mengimplementasikan pemilihan prompt berdasarkan mode Gemini dan pemrosesan embedding dokumen secara batch.

Author: chesterit21

sfcore-ai/crates/rag-api-server/config/settings.toml

@@ -77,6 +77,40 @@ acquire_timeout_ms = 60000
 embedding_batch_size = 10
 
 [prompts]
+
+[prompts.local]
+main_system_prompt = """
+Anda adalah asisten AI cerdas untuk Sistem Manajemen Dokumen.
+Waktu Server: {{CURRENT_DATETIME}}
+Dokumen Aktif:
+{{DOC_LIST}}
+
+Instruksi:
+- Jawab pertanyaan berdasarkan konteks dokumen.
+- Gunakan Bahasa Indonesia yang natural dan jelas.
+- Jika tidak ada info di dokumen, katakan jujur.
+
+Referensi:
+- [Judul Dokumen]
+"""
+
+context_extraction_system_prompt = """
+Konteks Dokumen:
+{{CHUNKS}}
+"""
+
+rag_query_system_prompt = """
+Anda adalah asisten AI RAG. Jawab berdasarkan konteks berikut:
+{{CONTEXT}}
+"""
+
+deep_scan_system_prompt = """
+User Query: "{{QUERY}}"
+Pilih chunk ID yang relevan dari daftar di bawah.
+Output JSON: {"relevant_chunk_ids": [1, 2, ...]}
+"""
+
+[prompts.gemini]
 main_system_prompt = """
 Anda adalah asisten AI cerdas untuk Sistem Manajemen Dokumen.
 Waktu Server: {{CURRENT_DATETIME}}

sfcore-ai/crates/rag-api-server/src/config/settings.rs

@@ -95,10 +95,16 @@ pub struct RagConfig {
 
 #[derive(Debug, Deserialize, Serialize, Clone)]
 pub struct PromptsConfig {
+    pub local: PromptSet,
+    pub gemini: PromptSet,
+}
+
+#[derive(Debug, Deserialize, Serialize, Clone)]
+pub struct PromptSet {
     pub main_system_prompt: String,
     pub context_extraction_system_prompt: String,
-    pub rag_query_system_prompt: String, // NEW
-    pub deep_scan_system_prompt: String, // NEW
+    pub rag_query_system_prompt: String,
+    pub deep_scan_system_prompt: String,
 }
 
 impl Settings {

sfcore-ai/crates/rag-api-server/src/main.rs

@@ -117,6 +117,17 @@ async fn main() -> Result<()> {
         }
     }
 
+    // Select Prompts based on Mode
+    let active_prompts = if let Some(gemini) = &settings.gemini {
+        if gemini.enabled {
+             settings.prompts.gemini.clone()
+        } else {
+             settings.prompts.local.clone()
+        }
+    } else {
+        settings.prompts.local.clone()
+    };
+
     // Initialize services
     let embedding_service = Arc::new(EmbeddingService::new(
         final_embedding_config.base_url.clone(),
@@ -127,7 +138,7 @@ async fn main() -> Result<()> {
 
     let llm_service = Arc::new(LlmService::new(
         final_llm_config.clone(),
-        settings.prompts.context_extraction_system_prompt.clone(),
+        active_prompts.context_extraction_system_prompt.clone(),
         limiters.clone(),
     ));
 
@@ -154,8 +165,8 @@ async fn main() -> Result<()> {
         Box::new((*llm_service).clone()),
         logger.clone(),
         settings.llm.stream_response,
-        settings.prompts.main_system_prompt.clone(),
-        settings.prompts.deep_scan_system_prompt.clone(),
+        active_prompts.main_system_prompt.clone(),
+        active_prompts.deep_scan_system_prompt.clone(),
         settings.rag.clone(),
     ));
     info!("✅ Conversation manager initialized");

sfcore-ai/crates/rag-api-server/src/services/document_service.rs

@@ -229,19 +229,37 @@ impl DocumentService {
             return Err(ApiError::BadRequest("Failed to create chunks".to_string()));
         }
 
+        // 4. Generate embeddings (batch) or Fallback
         // 4. Generate embeddings (batch) or Fallback
         report_progress(0.6, "Generating embeddings (this might take a while)...".to_string(), "embedding-inprogress".to_string());
         let texts: Vec<String> = chunks.clone();
 
-        let embeddings = match self.embedding_service.embed_batch(texts.clone()).await {
-            Ok(e) => e,
-            Err(err) => {
-                warn!("Embedding failed for document {} (falling back to zerovec): {}", document_id, err);
-                // Fallback to zero vectors so Deep Scan can still work
-                let dim = self.embedding_service.dimension;
-                vec![vec![0.0; dim]; texts.len()]
-            }
-        };
+        // Use configured batch size
+        let batch_size = self.embedding_batch_size.max(1);
+        let mut embeddings = Vec::with_capacity(texts.len());
+        let total_batches = (texts.len() + batch_size - 1) / batch_size;
+
+        for (i, batch_texts) in texts.chunks(batch_size).enumerate() {
+             // Report progress for this batch
+             report_progress(
+                 0.6 + (0.2 * (i as f64 / total_batches as f64)), 
+                 format!("Embedding batch {}/{}...", i + 1, total_batches), 
+                 "embedding-inprogress".to_string()
+             );
+
+             // Embed batch
+             match self.embedding_service.embed_batch(batch_texts.to_vec()).await {
+                Ok(batch_embs) => {
+                    embeddings.extend(batch_embs);
+                },
+                Err(err) => {
+                    warn!("Embedding failed for batch {}/{} of document {} (falling back to zerovec): {}", i + 1, total_batches, document_id, err);
+                    // Fallback to zero vectors so Deep Scan can still work for this batch
+                    let dim = self.embedding_service.dimension;
+                    embeddings.extend(vec![vec![0.0; dim]; batch_texts.len()]);
+                }
+             }
+        }
         // debug!("Generated {} embeddings", embeddings.len());
 
         // 5. Build chunk data