Recently, I built an end-to-end Agentic Conversational FAQ Assistant.

A few things worth sharing from the process:

• The agent can work on any dataset and maintains contextual awareness and memory while answering user queries.

• It retrieves information using an FAISS vector index built over the dataset.

• To separate the knowledge layer from the LLM layer, I designed the system so the index can be updated independently without changing the model itself.

• One of the main challenges in development was memory management. Multi-turn conversations become meaningless if the agent forgets what the user said two messages ago, but you can't send the full history on every call.

• To solve this, I used AWS Bedrock AgentCore’s memory layer, which scopes context per actor and session, allowing the agent to continue conversations without reloading everything.

• Orchestration is handled using LangGraph because the logic is not always linear. Depending on the retrieval results, the agent can dynamically choose different tool paths, which is difficult to manage cleanly using a simple linear chain.

• I also built a Streamlit-based interface to interact with the assistant and test multi-turn conversational behavior.

• For the knowledge base, I extracted and prepared the dataset from Jio's official website.

Code: Agentcore-Project