DocuTalk-AI is an intelligent retrieval-augmented generation (RAG) application powered by Google's Gemini 2.5 Flash.
Unlike basic chatbots, this system employs a ReAct (Reasoning + Acting) Agent architecture. It intelligently decides whether to answer questions based on your uploaded Knowledge Base (PDF, CSV, TXT, MD) or via Web Search (DuckDuckGo) if the document lacks specific information. It features incremental indexing, strict language control, and session management.
Powered by LangChain, the app creates a "Brain" that follows a strict priority protocol:
- Priority 1: Check Uploaded Documents (Vector Search via FAISS).
- Priority 2: Use Internal Knowledge.
- Priority 3: Web Search (Fallback if info is missing).
- Smart Processing: The app tracks filenames (
processed_filesstate). If you add a new file to an existing batch, it only processes the new file without re-embedding the old ones. - Optimization: Saves time and API quota by appending to the existing FAISS vector store instead of rebuilding it from scratch.
Features a custom "Translator-Researcher" prompt injection.
- User Control: Select English or Indonesian in the sidebar.
- Behavior: The AI ignores the source document's language and forces the final output into your selected language (e.g., querying an English Journal but getting the answer in Indonesian).
- Chat History Export: Download your entire conversation analysis as a JSON file.
- Memory Handling: Uses
ConversationSummaryMemoryto maintain context without overloading the token limit. - Safety Rails: Robust error handling for API Quotas (429), Invalid Keys, and Empty/Corrupt Files.
- Creativity Slider: Adjust the
temperature(0.0 - 1.0) to switch between strict fact-extraction and creative summarization. - Visual Thinking: Real-time visibility into the Agent's thought process (Observation -> Thought -> Action) via
StreamlitCallbackHandler.
- LLM: Google Gemini 2.5 Flash.
- Framework: Streamlit.
- Orchestration: LangChain (Agents, Tools, Chains).
- Vector Database: FAISS (In-memory).
- Embeddings: GoogleGenerativeAIEmbeddings (
models/gemini-embedding-001). - Search Tool: DuckDuckGo Search Run.
-
Clone the Repository
git clone https://github.com/viochris/DocuTalk-AI cd DocuTalk-AI -
Install Dependencies
pip install -r requirements.txt
-
Run the Application
streamlit run file.py
- Setup:
- Get your API Key from Google AI Studio.
- Enter the key in the sidebar.
- Configuration:
- Choose Response Language (English/Indonesian).
- Adjust Creativity Level.
- Build Knowledge Base:
- Upload your documents (PDF, CSV, MD, or TXT).
- Click "🚀 Process & Embed Documents".
- Wait for the "Successfully added" toast notification.
- Chat & Analyze:
- Ask questions like "Summarize the methodology" or "Analyze the data trends in the CSV".
- Use the Recommended Questions expander for quick start ideas.
- Export:
- Click "📥 Download Chat History" to save your insights.
- Session Volatility: Since FAISS is stored in RAM (Session State), refreshing the browser will clear the document index.
- API Quotas: Heavy usage of the Embedding model or LLM may hit Google's free tier rate limits (captured by the app's error handling).
The clean, user-friendly landing page welcoming users to DocuTalk-AI. It features a clear call-to-action for uploading documents and initializing the secure environment.
The remaining part of the sidebar configuration that is not visible in the Home UI screenshot.
The core experience in action. Once documents are processed, the chat interface enables natural language queries, displaying precise answers derived directly from the uploaded content.
Author: Silvio Christian, Joe "Unlock insights from your documents instantly."