Context-Aware Chatbot Optimization
Budget: $30 – $250 USD
I need the current RAG-based chatbot tightened up so it answers customer service issues with context-driven accuracy. When a visitor greets the bot or slips into small talk, the response must remain professional and formal, then transition smoothly back to the user’s request. After each reply, the bot should pose one concise follow-up question that relates directly to the user’s previous inquiry, helping close any knowledge gaps.
Memory is critical. The bot should keep only the most recent question-and-answer pair verbatim plus a compact summary of the five prior exchanges; everything beyond that can be forgotten. I will share the existing Python code (LangChain + OpenAI) and access to our knowledge base so you can refine prompts, retrieval logic, and embeddings.
Deliverables
• Updated source code with comments explaining retrieval tweaks, prompt changes, and the exact memory implementation
• A brief README detailing setup, memory flow, and how to refresh embeddings
• Short demo (notebook or video) showing:
1. Professional greeting and small talk handling
2. Accurate answer to a customer service issue drawn from context
3. A follow-up question generated from that same interaction
4. Proper operation of the six-turn memory window
Acceptance criteria: answers match documented truth within 1 % error, greetings stay formal, follow-up questions are relevant, and the memory behaves exactly as described.
Memory is critical. The bot should keep only the most recent question-and-answer pair verbatim plus a compact summary of the five prior exchanges; everything beyond that can be forgotten. I will share the existing Python code (LangChain + OpenAI) and access to our knowledge base so you can refine prompts, retrieval logic, and embeddings.
Deliverables
• Updated source code with comments explaining retrieval tweaks, prompt changes, and the exact memory implementation
• A brief README detailing setup, memory flow, and how to refresh embeddings
• Short demo (notebook or video) showing:
1. Professional greeting and small talk handling
2. Accurate answer to a customer service issue drawn from context
3. A follow-up question generated from that same interaction
4. Proper operation of the six-turn memory window
Acceptance criteria: answers match documented truth within 1 % error, greetings stay formal, follow-up questions are relevant, and the memory behaves exactly as described.