Remove ngrok.exe from tracking

Dockerfile

@@ -0,0 +1,20 @@
+FROM python:3.13-slim
+
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y gcc g++ curl git && rm -rf /var/lib/apt/lists/*
+
+COPY requirements-deploy.txt ./
+RUN pip install --no-cache-dir -r requirements-deploy.txt
+
+COPY src/ ./src/
+COPY app/ ./app/
+COPY .chainlit/ ./.chainlit/
+COPY chainlit.md ./chainlit.md
+
+RUN useradd -m -u 1000 user && chown -R user:user /app
+USER user
+
+EXPOSE 7860
+
+CMD ["chainlit", "run", "app/vilexagent_ui.py", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -0,0 +1,210 @@
+# ⚖️ ViLexAgent
+
+> **A Multi-Agent RAG System for Vietnamese Legal Q&A**  
+> Supports Vietnamese labor law, food safety regulations, and international trade agreements (EVFTA, CPTPP).
+
+---
+
+## Overview
+
+ViLexAgent is an agentic Retrieval-Augmented Generation (RAG) system that answers complex legal questions about Vietnamese law and its compliance with international trade agreements. The system decomposes user queries into sub-questions, retrieves relevant legal documents from a vector database, cross-references domestic law with international standards, and synthesizes a cited, legally-grounded answer.
+
+**Key capabilities:**
+- Query decomposition into domain-specific sub-questions (labor law, food safety)
+- Hybrid retrieval from domestic Vietnamese legal documents and international treaty clauses (EVFTA, CPTPP)
+- Automatic cross-reference analysis with alignment scoring (`aligned` / `conflict` / `gap` / `no_international`)
+- Expired document detection with explicit warnings
+- Step-by-step reasoning trace visible in the UI
+
+---
+
+## Architecture
+
+```
+User Query
+    │
+    ▼
+┌─────────────────────┐
+│   Query Decomposer  │  → Sub-questions + domain tags (labor / food_safety)
+└─────────────────────┘
+    │
+    ├──────────────────────────────────┐
+    ▼                                  ▼
+┌──────────────────┐        ┌──────────────────────────┐
+│ Domestic         │        │ International             │
+│ Retriever        │        │ Retriever                 │
+│ (Qdrant + Jina)  │        │ (Qdrant + Jina)           │
+└──────────────────┘        └──────────────────────────┘
+    │                                  │
+    └──────────────┬───────────────────┘
+                   ▼
+        ┌─────────────────────┐
+        │   Cross-Reference   │  → alignment: aligned / conflict / gap
+        └─────────────────────┘
+                   │
+                   ▼
+        ┌─────────────────────┐
+        │     Synthesizer     │  → Final answer with citations
+        └─────────────────────┘
+```
+
+**Tech Stack:**
+
+| Layer | Technology |
+|---|---|
+| UI | Chainlit 2.11 |
+| Agent Orchestration | LangGraph 1.1 |
+| Vector Database | Qdrant |
+| Embedding Model | Jina Embeddings v5 Small (4-bit, CUDA) |
+| LLM Backend | LiteLLM (OpenAI-compatible) |
+| OCR / PDF Parsing | PaddleOCR, PyMuPDF |
+| Vietnamese NLP | Underthesea |
+| Experiment Tracking | MLflow |
+| Evaluation | RAGAS |
+| Runtime | Python 3.13, Poetry |
+
+---
+
+## Project Structure
+
+```
+vilexagent/
+├── app/
+│   └── vilexagent_ui.py       # Chainlit UI entry point
+├── src/
+│   ├── agents/
+│   │   ├── state.py               # AgentState (LangGraph TypedDict)
+│   │   ├── query_decomposer.py    # Query decomposition node
+│   │   ├── domestic_retriever.py  # Vietnamese law retrieval node
+│   │   ├── international_retriever.py  # EVFTA/CPTPP retrieval node
+│   │   ├── cross_reference.py     # Cross-reference analysis node
+│   │   ├── synthesizer.py         # Answer synthesis node
+│   │   └── graph.py               # LangGraph pipeline definition
+│   ├── ingestion/                 # Document ingestion pipeline
+│   ├── retrieval/                 # Retrieval utilities
+│   └── utils/
+│       ├── llm.py                 # LLM client (LiteLLM wrapper)
+│       ├── model_loader.py        # Singleton embedding model loader
+│       ├── logger.py              # Loguru logger
+│       └── json_utils.py          # Robust JSON extractor
+├── evaluation/                    # RAGAS benchmark suite
+├── docker/                        # Docker/Qdrant setup
+├── tests/
+├── pyproject.toml
+└── chainlit.md
+```
+
+---
+
+## Installation
+
+### Prerequisites
+
+- Python 3.13
+- [Poetry](https://python-poetry.org/docs/#installation)
+- CUDA-capable GPU (recommended: 4GB+ VRAM)
+- [Qdrant](https://qdrant.tech/documentation/quick-start/) running locally on port `6333`
+- An OpenAI-compatible LLM API endpoint on port `3001`
+
+### Steps
+
+**1. Clone the repository**
+```bash
+git clone https://github.com/dnAnh1523/vilexagent.git
+cd vilexagent
+```
+
+**2. Install dependencies**
+```bash
+poetry install
+```
+
+**3. Set up environment variables**
+
+Create a `.env` file in the project root:
+```env
+LLM_BASE_URL=http://localhost:3001/v1
+LLM_API_KEY=your_api_key_here
+LLM_MODEL=your_model_name
+```
+
+**4. Start Qdrant**
+```bash
+docker compose -f docker/docker-compose.yml up -d
+```
+
+**5. Ingest documents**
+
+Place your Vietnamese legal PDFs and international treaty documents in the `data/` directory, then run the ingestion pipeline (see `src/ingestion/`).
+
+**6. Run the app**
+```bash
+poetry run chainlit run app/vilexagent_ui.py --port 8000
+```
+
+Open `http://localhost:8000` in your browser.
+
+---
+
+## Usage
+
+Once running, you can ask questions in Vietnamese about:
+
+- **Labor law** — probationary periods, wrongful termination compensation, minimum wage, overtime
+- **Food safety** — export conditions, hygiene standards, quarantine requirements
+- **International compliance** — whether Vietnamese law meets EVFTA/CPTPP standards
+
+The UI displays a step-by-step reasoning trace (collapsible) showing which documents were retrieved, cross-reference results, and alignment scores before the final answer.
+
+**Example questions:**
+```
+Thời gian thử việc tối đa theo pháp luật lao động Việt Nam là bao lâu?
+
+Nếu người sử dụng lao động đơn phương chấm dứt hợp đồng trái pháp luật
+thì phải bồi thường những gì cho người lao động?
+
+Việt Nam có đáp ứng các tiêu chuẩn lao động của CPTPP về tự do hiệp hội không?
+```
+
+---
+
+## Configuration
+
+Key settings in `chainlit.md` (UI welcome screen) and `.chainlit/config.toml`:
+
+| Setting | Location | Description |
+|---|---|---|
+| `cot` | `config.toml [UI]` | Chain-of-thought display: `"full"` shows all steps |
+| `session_timeout` | `config.toml [project]` | Session retention in seconds |
+| `LLM_BASE_URL` | `.env` | OpenAI-compatible LLM endpoint |
+| `LLM_MODEL` | `.env` | Model name passed to LiteLLM |
+
+---
+
+## Evaluation
+
+The system includes a RAGAS-based evaluation suite in `evaluation/` with a benchmark of labeled legal Q&A pairs across three difficulty types:
+
+- **Type A** — Single-domain factual queries (domestic law only)
+- **Type B** — Multi-aspect domestic queries requiring reasoning across clauses
+- **Type C** — Cross-reference queries requiring domestic + international alignment
+
+Run evaluation:
+```bash
+poetry run python evaluation/run_evaluation.py
+```
+
+Results are tracked with MLflow. Start the MLflow UI:
+```bash
+poetry run mlflow ui
+```
+
+---
+
+## License
+
+This project is for educational and portfolio purposes.
+
+---
+
+*Built with LangGraph · Qdrant · Chainlit · Jina Embeddings*

app/vilexagent_ui.py

@@ -1,5 +1,6 @@
 # app/vilexagent_ui.py
 import sys
+import traceback
 import chainlit as cl
 from dotenv import load_dotenv
 sys.path.append(r"E:\\vilexagent")
@@ -47,6 +48,7 @@ async def on_faq(action: cl.Action):
     query = action.payload.get("query", "")
     await on_message(cl.Message(content=query))
 
+
 @cl.on_message
 async def on_message(message: cl.Message):
     query = message.content.strip()
@@ -66,116 +68,128 @@ async def on_message(message: cl.Message):
         "error": None
     }
  
-    async with cl.Step(name="💭 Thought", type="run") as thought:
-        thought.input = query
- 
-        # --- Step 1: Query Decomposition ---
-        async with cl.Step(name="🔍 Phân tích câu hỏi", type="tool") as step1:
-            step1.input = query
-            from src.agents.query_decomposer import query_decomposer_node
-            decomp_result = query_decomposer_node(state)
-            state.update(decomp_result)
-            sub_questions = state.get("sub_questions", [])
-            requires_intl = state.get("requires_international", False)
- 
-            sub_q_text = "\n".join([
-                f"- [{sq['source']}|{sq['domain']}] {sq['question']}"
-                for sq in sub_questions
-            ])
-            step1.output = (
-                f"Chia thành **{len(sub_questions)}** câu hỏi con:\n{sub_q_text}\n"
-                f"Cần tra cứu quốc tế: {'✅' if requires_intl else '❌'}"
-            )
- 
-        # --- Step 2: Domestic Retrieval ---
-        async with cl.Step(name="📚 Tra cứu pháp luật Việt Nam", type="tool") as step2:
-            step2.input = f"{len([sq for sq in sub_questions if sq['source'] in ('domestic','both')])} câu hỏi nội địa"
-            from src.agents.domestic_retriever import domestic_retriever_node
-            domestic_result = domestic_retriever_node(state)
-            state.update(domestic_result)
-            domestic_chunks = state.get("domestic_chunks", [])
- 
-            if domestic_chunks:
-                docs_text = "\n".join([
-                    f"- {c['title'][:60]} — {c['article_number']} "
-                    f"({'⚠️ hết hiệu lực' if 'hết hiệu lực' in c.get('tinh_trang_hieu_luc','').lower() else '✅ còn hiệu lực'})"
-                    for c in domestic_chunks[:5]
+    try:
+        async with cl.Step(name="Đang suy nghĩ...", type="run") as thought:
+            thought.input = query
+    
+            # --- Step 1: Query Decomposition ---
+            async with cl.Step(name="🔍 Phân tích câu hỏi", type="tool") as step1:
+                step1.input = query
+                from src.agents.query_decomposer import query_decomposer_node
+                # FIX: Bọc make_async để không block UI
+                decomp_result = await cl.make_async(query_decomposer_node)(state)
+                state.update(decomp_result)
+                sub_questions = state.get("sub_questions", [])
+                requires_intl = state.get("requires_international", False)
+    
+                sub_q_text = "\n".join([
+                    f"- [{sq['source']}|{sq['domain']}] {sq['question']}"
+                    for sq in sub_questions
                 ])
-                step2.output = f"Tìm thấy **{len(domestic_chunks)}** văn bản:\n{docs_text}"
-            else:
-                step2.output = "Không tìm thấy văn bản nội địa liên quan."
- 
-        # --- Step 3: International Retrieval (if needed) ---
-        if requires_intl:
-            async with cl.Step(name="🌐 Tra cứu tiêu chuẩn quốc tế", type="tool") as step3:
-                step3.input = f"{len([sq for sq in sub_questions if sq['source'] in ('international','both')])} câu hỏi quốc tế"
-                from src.agents.international_retriever import international_retriever_node
-                intl_result = international_retriever_node(state)
-                state.update(intl_result)
-                intl_chunks = state.get("international_chunks", [])
- 
-                if intl_chunks:
-                    intl_text = "\n".join([
-                        f"- [{c.get('agreement','')}] {c['title'][:60]} — {c['article_number']}"
-                        for c in intl_chunks[:5]
+                step1.output = (
+                    f"Chia thành **{len(sub_questions)}** câu hỏi con:\n{sub_q_text}\n"
+                    f"Cần tra cứu quốc tế: {'✅' if requires_intl else '❌'}"
+                )
+    
+            # --- Step 2: Domestic Retrieval ---
+            async with cl.Step(name="📚 Tra cứu pháp luật Việt Nam", type="tool") as step2:
+                step2.input = f"{len([sq for sq in sub_questions if sq['source'] in ('domestic','both')])} câu hỏi nội địa"
+                from src.agents.domestic_retriever import domestic_retriever_node
+                # FIX: Bọc make_async
+                domestic_result = await cl.make_async(domestic_retriever_node)(state)
+                state.update(domestic_result)
+                domestic_chunks = state.get("domestic_chunks", [])
+    
+                if domestic_chunks:
+                    docs_text = "\n".join([
+                        f"- {c['title'][:60]} — {c['article_number']} "
+                        # Thêm ép kiểu str() để chống lỗi NoneType lặt vặt
+                        f"({'⚠️ hết hiệu lực' if 'hết hiệu lực' in str(c.get('tinh_trang_hieu_luc', '')).lower() else '✅ còn hiệu lực'})"
+                        for c in domestic_chunks[:5]
                     ])
-                    step3.output = f"Tìm thấy **{len(intl_chunks)}** điều khoản quốc tế:\n{intl_text}"
+                    step2.output = f"Tìm thấy **{len(domestic_chunks)}** văn bản:\n{docs_text}"
                 else:
-                    step3.output = "Không tìm thấy điều khoản quốc tế liên quan."
- 
-        # --- Step 4: Cross-Reference --- luôn chạy
-        async with cl.Step(name="⚖️ Đối chiếu pháp luật", type="tool") as step4:
-            step4.input = (
-                f"{len(state.get('domestic_chunks', []))} văn bản nội địa × "
-                f"{len(state.get('international_chunks', []))} điều khoản quốc tế"
+                    step2.output = "Không tìm thấy văn bản nội địa liên quan."
+    
+            # --- Step 3: International Retrieval (if needed) ---
+            if requires_intl:
+                async with cl.Step(name="🌐 Tra cứu tiêu chuẩn quốc tế", type="tool") as step3:
+                    step3.input = f"{len([sq for sq in sub_questions if sq['source'] in ('international','both')])} câu hỏi quốc tế"
+                    from src.agents.international_retriever import international_retriever_node
+                    # FIX: Bọc make_async
+                    intl_result = await cl.make_async(international_retriever_node)(state)
+                    state.update(intl_result)
+                    intl_chunks = state.get("international_chunks", [])
+    
+                    if intl_chunks:
+                        intl_text = "\n".join([
+                            f"- [{c.get('agreement','')}] {c['title'][:60]} — {c['article_number']}"
+                            for c in intl_chunks[:5]
+                        ])
+                        step3.output = f"Tìm thấy **{len(intl_chunks)}** điều khoản quốc tế:\n{intl_text}"
+                    else:
+                        step3.output = "Không tìm thấy điều khoản quốc tế liên quan."
+    
+            # --- Step 4: Cross-Reference ---
+            async with cl.Step(name="⚖️ Đối chiếu pháp luật", type="tool") as step4:
+                step4.input = (
+                    f"{len(state.get('domestic_chunks', []))} văn bản nội địa × "
+                    f"{len(state.get('international_chunks', []))} điều khoản quốc tế"
+                )
+                from src.agents.cross_reference import cross_reference_node
+                # FIX: Bọc make_async
+                xref_result = await cl.make_async(cross_reference_node)(state)
+                state.update(xref_result)
+                cross_ref = state.get("cross_reference") or {}
+    
+                alignment = cross_ref.get("alignment", "unknown")
+                alignment_emoji = {
+                    "aligned": "✅ Phù hợp",
+                    "conflict": "❌ Mâu thuẫn",
+                    "gap": "⚠️ Còn khoảng cách",
+                    "no_international": "ℹ️ Không áp dụng"
+                }.get(alignment, alignment)
+    
+                step4.output = (
+                    f"Kết quả đối chiếu: **{alignment_emoji}**\n\n"
+                    f"{str(cross_ref.get('explanation', ''))[:300]}"
+                )
+    
+            # --- Step 5: Synthesis ---
+            async with cl.Step(name="✍️ Tổng hợp câu trả lời", type="tool") as step5:
+                step5.input = "Tổng hợp từ tất cả nguồn"
+                from src.agents.synthesizer import synthesizer_node
+                # FIX: Bọc make_async
+                synth_result = await cl.make_async(synthesizer_node)(state)
+                state.update(synth_result)
+                step5.output = f"Đã tạo câu trả lời ({len(state.get('final_answer',''))} ký tự)"
+    
+            thought.output = (
+                f"{len(state.get('domestic_chunks', []))} văn bản nội địa · "
+                f"{len(state.get('international_chunks', []))} điều khoản quốc tế · "
+                f"{len(state.get('final_answer',''))} ký tự"
             )
-            from src.agents.cross_reference import cross_reference_node
-            xref_result = cross_reference_node(state)
-            state.update(xref_result)
-            cross_ref = state.get("cross_reference") or {}
- 
-            alignment = cross_ref.get("alignment", "unknown")
-            alignment_emoji = {
-                "aligned": "✅ Phù hợp",
-                "conflict": "❌ Mâu thuẫn",
-                "gap": "⚠️ Còn khoảng cách",
-                "no_international": "ℹ️ Không áp dụng"
-            }.get(alignment, alignment)
- 
-            step4.output = (
-                f"Kết quả đối chiếu: **{alignment_emoji}**\n\n"
-                f"{cross_ref.get('explanation', '')[:300]}"
+    
+        # --- Final Answer ---
+        final_answer = state.get("final_answer", "Xin lỗi, tôi không thể trả lời câu hỏi này.")
+        citations = state.get("citations", [])
+        has_expired = state.get("has_expired_docs", False)
+    
+        footer = ""
+        if citations:
+            footer += "\n\n---\n**📌 Nguồn tham khảo:**\n"
+            for c in citations[:8]:
+                footer += f"- {c}\n"
+    
+        if has_expired:
+            footer += (
+                "\n\n> ⚠️ **Lưu ý:** Một số văn bản pháp luật trong kết quả tra cứu "
+                "đã **hết hiệu lực**. Vui lòng chỉ sử dụng văn bản còn hiệu lực làm căn cứ pháp lý."
             )
- 
-        # --- Step 5: Synthesis ---
-        async with cl.Step(name="✍️ Tổng hợp câu trả lời", type="tool") as step5:
-            step5.input = "Tổng hợp từ tất cả nguồn"
-            from src.agents.synthesizer import synthesizer_node
-            synth_result = synthesizer_node(state)
-            state.update(synth_result)
-            step5.output = f"Đã tạo câu trả lời ({len(state.get('final_answer',''))} ký tự)"
- 
-        thought.output = (
-            f"{len(state.get('domestic_chunks', []))} văn bản nội địa · "
-            f"{len(state.get('international_chunks', []))} điều khoản quốc tế · "
-            f"{len(state.get('final_answer',''))} ký tự"
-        )
- 
-    # --- Final Answer ---
-    final_answer = state.get("final_answer", "Xin lỗi, tôi không thể trả lời câu hỏi này.")
-    citations = state.get("citations", [])
-    has_expired = state.get("has_expired_docs", False)
- 
-    footer = ""
-    if citations:
-        footer += "\n\n---\n**📌 Nguồn tham khảo:**\n"
-        for c in citations[:8]:
-            footer += f"- {c}\n"
- 
-    if has_expired:
-        footer += (
-            "\n\n> ⚠️ **Lưu ý:** Một số văn bản pháp luật trong kết quả tra cứu "
-            "đã **hết hiệu lực**. Vui lòng chỉ sử dụng văn bản còn hiệu lực làm căn cứ pháp lý."
-        )
- 
-    await cl.Message(content=final_answer + footer).send()
+    
+        await cl.Message(content=final_answer + footer).send()
+
+    except Exception as e:
+        # Bắt mọi lỗi xảy ra trong UI và báo thẳng ra màn hình
+        error_msg = f"❌ **Lỗi Hệ Thống UI:**\n```python\n{traceback.format_exc()}\n```"
+        await cl.Message(content=error_msg).send()

pyproject.toml

@@ -14,7 +14,6 @@ dependencies = [
     "langchain-community (>=0.4.1,<0.5.0)",
     "qdrant-client (>=1.17.1,<2.0.0)",
     "sentence-transformers (>=5.4.1,<6.0.0)",
-    "ragas (>=0.4.3,<0.5.0)",
     "mlflow (>=3.11.1,<4.0.0)",
     "pymupdf (>=1.27.2.3,<2.0.0.0)",
     "beautifulsoup4 (>=4.14.3,<5.0.0)",
@@ -30,10 +29,7 @@ dependencies = [
     "peft (>=0.19.1,<0.20.0)",
     "torch (>=2.11.0,<3.0.0)",
     "accelerate (>=1.13.0,<2.0.0)",
-    "bitsandbytes (>=0.49.2,<0.50.0)",
     "google-generativeai (>=0.8.6,<0.9.0)",
-    "litellm (>=1.83.14,<2.0.0)",
-    "langchain-ollama (>=1.1.0,<2.0.0)",
     "chainlit (>=2.11.1,<3.0.0)"
 ]
 

requirements-deploy.txt

@@ -0,0 +1,14 @@
+langgraph>=1.1.10
+langchain-google-genai>=4.2.2
+langchain-community>=0.4.1
+langchain-openai>=0.3.0
+qdrant-client>=1.17.1
+sentence-transformers>=5.4.1
+torch --index-url https://download.pytorch.org/whl/cpu
+peft>=0.19.1
+accelerate>=1.13.0
+huggingface-hub>=1.13.0
+python-dotenv>=1.2.2
+pydantic>=2.12.5
+loguru>=0.7.3
+chainlit>=2.11.1

src/agents/domestic_retriever.py

@@ -1,6 +1,5 @@
 # src/agents/domestic_retriever.py
-import os
-from qdrant_client import QdrantClient
+from src.utils.qdrant_client import get_qdrant_client
 from qdrant_client.models import Filter, FieldCondition, MatchValue
 from src.agents.state import AgentState
 from src.utils.logger import logger
@@ -19,7 +18,7 @@ _client = None
 def get_client():
     global _client
     if _client is None:
-        _client = QdrantClient(url=os.getenv("QDRANT_URL", "http://localhost:6333"))
+        _client = get_qdrant_client()
     return _client
 
 def retrieve_domestic(query: str, domain: str) -> list[dict]:

src/agents/international_retriever.py

@@ -1,6 +1,5 @@
 # src/agents/international_retriever.py
-import os
-from qdrant_client import QdrantClient
+from src.utils.qdrant_client import get_qdrant_client
 from qdrant_client.models import Filter, FieldCondition, MatchValue
 from src.agents.state import AgentState
 from src.utils.logger import logger
@@ -18,7 +17,7 @@ _client = None
 def get_client():
     global _client
     if _client is None:
-        _client = QdrantClient(url=os.getenv("QDRANT_URL", "http://localhost:6333"))
+        _client = get_qdrant_client()
     return _client
 
 def retrieve_international(query: str, domain: str) -> list[dict]:

src/retrieval/baseline.py

@@ -1,9 +1,8 @@
 # src/retrieval/baseline.py
-import os
-from qdrant_client import QdrantClient
 from qdrant_client.models import Filter, FieldCondition, MatchValue
 from src.utils.model_loader import get_embedding_model
 from src.utils.logger import logger
+from src.utils.qdrant_client import get_qdrant_client
 
 DOMESTIC_COLLECTION = "vilexagent_domestic"
 INTERNATIONAL_COLLECTION = "vilexagent_international"
@@ -12,7 +11,7 @@ TOP_K = 5
 class BaselineRetriever:
     def __init__(self):
         self.model = get_embedding_model()
-        self.client = QdrantClient(url=os.getenv("QDRANT_URL", "http://localhost:6333"))
+        self.client = get_qdrant_client()
         logger.success("BaselineRetriever ready")
 
     def retrieve(self, query: str, source: str = "domestic", domain: str = None, top_k: int = TOP_K) -> list[dict]:

src/utils/llm.py

@@ -1,29 +1,40 @@
 # src/utils/llm.py
 import os
 from dotenv import load_dotenv
-from langchain_openai import ChatOpenAI
 from src.utils.logger import logger
 
 load_dotenv()
 
-FREELLM_BASE_URL = os.getenv("FREELLM_BASE_URL", "http://localhost:3001/v1")
-FREELLM_API_KEY = os.getenv("FREELLM_API_KEY")
-
-if not FREELLM_API_KEY:
-    raise ValueError("FREELLM_API_KEY not set in .env")
-
 _llm = None
 
-def get_llm(temperature: float = 0) -> ChatOpenAI:
+
+def get_llm(temperature: float = 0):
     global _llm
     if _llm is None:
-        logger.info(f"Initializing LLM via FreeLLMAPI at {FREELLM_BASE_URL}")
-        _llm = ChatOpenAI(
-            model="meta-llama/llama-4-scout-17b-16e-instruct",
-            base_url=FREELLM_BASE_URL,
-            api_key=FREELLM_API_KEY,
-            temperature=temperature,
-            max_tokens=4096,
-        )
+        freellm_key = os.getenv("FREELLM_API_KEY")
+        freellm_url = os.getenv("FREELLM_BASE_URL")
+
+        if freellm_key and freellm_url:
+            # Local: use FreeLLMAPI proxy
+            from langchain_openai import ChatOpenAI
+            logger.info(f"Initializing LLM via FreeLLMAPI at {freellm_url}")
+            _llm = ChatOpenAI(
+                model="meta-llama/llama-4-scout-17b-16e-instruct",
+                base_url=freellm_url,
+                api_key=freellm_key,
+                temperature=temperature,
+                max_tokens=4096,
+            )
+        else:
+            # Cloud: use Gemini directly
+            from langchain_google_genai import ChatGoogleGenerativeAI
+            logger.info("Initializing LLM via Gemini API (cloud mode)")
+            _llm = ChatGoogleGenerativeAI(
+                model="gemini-2.5-flash",
+                google_api_key=os.getenv("GOOGLE_API_KEY"),
+                temperature=temperature,
+                max_tokens=4096,
+            )
+
         logger.success("LLM ready")
     return _llm

src/utils/migrate_to_cloud.py

@@ -0,0 +1,121 @@
+# src/utils/migrate_to_cloud.py
+"""
+Migrates vectors from local Qdrant to Qdrant Cloud.
+Reads all points from local collections and upserts to cloud.
+"""
+import os
+import time
+from dotenv import load_dotenv
+from qdrant_client import QdrantClient
+from qdrant_client.models import VectorParams, Distance, PayloadSchemaType, PointStruct
+from src.utils.logger import logger
+
+load_dotenv()
+
+LOCAL_URL = os.getenv("QDRANT_URL", "http://localhost:6333")
+CLOUD_URL = os.getenv("QDRANT_CLOUD_URL")
+CLOUD_API_KEY = os.getenv("QDRANT_API_KEY")
+COLLECTIONS = ["vilexagent_domestic", "vilexagent_international"]
+
+# Reduced batch size specifically for Qdrant Cloud Free Tier
+BATCH_SIZE = 25
+
+
+def migrate():
+    if not CLOUD_URL or not CLOUD_API_KEY:
+        raise ValueError("QDRANT_CLOUD_URL and QDRANT_API_KEY must be set in .env")
+
+    logger.info(f"Connecting to local Qdrant at {LOCAL_URL}...")
+    local = QdrantClient(url=LOCAL_URL)
+
+    # Increased timeout to 120 seconds to prevent WriteTimeout on large payloads
+    logger.info(f"Connecting to Qdrant Cloud at {CLOUD_URL}...")
+    cloud = QdrantClient(url=CLOUD_URL, api_key=CLOUD_API_KEY, timeout=120.0)
+
+    for collection_name in COLLECTIONS:
+        logger.info(f"\n{'='*50}")
+        logger.info(f"Migrating collection: {collection_name}")
+
+        # Get local collection info
+        try:
+            local_info = local.get_collection(collection_name)
+        except Exception as e:
+            logger.warning(f"Collection {collection_name} not found locally: {e}")
+            continue
+
+        vector_size = local_info.config.params.vectors.size
+        local_count = local.count(collection_name).count
+        logger.info(f"Local vectors: {local_count}, dimension: {vector_size}")
+
+        # Create cloud collection if not exists
+        existing = [c.name for c in cloud.get_collections().collections]
+        if collection_name in existing:
+            cloud_count = cloud.count(collection_name).count
+            logger.warning(f"Collection already exists in cloud with {cloud_count} vectors")
+            if cloud_count == local_count:
+                logger.success(f"Already synced, skipping")
+                continue
+            logger.info("Deleting and recreating...")
+            cloud.delete_collection(collection_name)
+
+        cloud.create_collection(
+            collection_name=collection_name,
+            vectors_config=VectorParams(size=vector_size, distance=Distance.COSINE)
+        )
+
+        for field in ["domain", "source", "language", "tinh_trang_hieu_luc", "loai_van_ban"]:
+            cloud.create_payload_index(
+                collection_name=collection_name,
+                field_name=field,
+                field_schema=PayloadSchemaType.KEYWORD
+            )
+
+        logger.success(f"Cloud collection created (dim={vector_size})")
+
+        # Migrate in batches using scroll
+        offset = None
+        total_migrated = 0
+
+        while True:
+            results, next_offset = local.scroll(
+                collection_name=collection_name,
+                limit=BATCH_SIZE,
+                offset=offset,
+                with_vectors=True,
+                with_payload=True
+            )
+
+            if not results:
+                break
+
+            # Convert Record objects to PointStruct objects
+            points_to_upsert = [
+                PointStruct(
+                    id=record.id,
+                    vector=record.vector,
+                    payload=record.payload
+                ) for record in results
+            ]
+
+            cloud.upsert(
+                collection_name=collection_name,
+                points=points_to_upsert
+            )
+
+            total_migrated += len(results)
+            logger.info(f"  Migrated {total_migrated}/{local_count}...")
+
+            # Add a pause to prevent overwhelming the Free Tier CPU
+            time.sleep(0.5) 
+
+            if next_offset is None:
+                break
+            offset = next_offset
+
+        cloud_count = cloud.count(collection_name).count
+        logger.success(f"Migration complete: {cloud_count} vectors in cloud")
+
+    logger.success("\nAll collections migrated successfully.")
+
+if __name__ == "__main__":
+    migrate()

src/utils/model_loader.py

@@ -20,8 +20,7 @@ def get_embedding_model():
             device=device,
             trust_remote_code=True,
             model_kwargs={
-                "torch_dtype": torch.bfloat16, 
-                "load_in_4bit": True,
+                "torch_dtype": torch.float16 if device == "cuda" else torch.float32,
                 "default_task": "retrieval"
             }
         )

src/utils/qdrant_client.py

@@ -0,0 +1,20 @@
+# src/utils/qdrant_client.py
+import os
+from qdrant_client import QdrantClient
+from src.utils.logger import logger
+
+_client = None
+
+def get_qdrant_client() -> QdrantClient:
+    global _client
+    if _client is None:
+        api_key = os.getenv("QDRANT_API_KEY")
+        if api_key:
+            url = os.getenv("QDRANT_CLOUD_URL")
+            logger.info(f"Connecting to Qdrant Cloud at {url}")
+            _client = QdrantClient(url=url, api_key=api_key)
+        else:
+            url = os.getenv("QDRANT_URL", "http://localhost:6333")
+            logger.info(f"Connecting to local Qdrant at {url}")
+            _client = QdrantClient(url=url)
+    return _client