FINCA Chatbot

-
Created a chatbot to query the internal document store


Project Goals

FINCA aimed to create a chatbot to efficiently query and retrieve information from their internal document store, enhancing the accessibility and usability of past data. For this project, I worked as the network architect and deployment automation designer, creating all the terraform templates, deploying AWS Bedrock Knowledge Bases, and ensuring that security was maintained throughout the environment.

Project Deliverables

  • Custom chatbot agent using AWS Lambda, Bedrock Knowledge Bases, Langchain, and Langraph
  • Integration with Microsoft Entra for secure authentication and authorization
  • User-friendly frontend interface for seamless interaction
  • DynamoDB for chat storage and Cohere’s embedding model for vector DB in OpenSearch
  • Deployment using Terraform for easy replication across environments

Challenges Encountered

The main challenge was handling the diverse formats of data within the document store, including text, JSON, CSV, and mixed-media. Additionally, the chatbot struggled with providing summary numbers for tabular data, as it required additional tooling beyond the project’s scope.

Project Outcomes

The chatbot successfully enabled users to retrieve relevant information from the old document stores, significantly improving their usability. The integration with Microsoft Entra ensured secure access to data sources. Despite limitations in handling tabular data summaries, the overall project outcome was positive, with the chatbot providing reasonable information to client queries.

Key Takeaways

  • Integrating various technologies (AWS Lambda, Bedrock Knowledge Bases, Langchain, Langraph) can create powerful chatbot solutions
  • Centralized authentication and authorization (Microsoft Entra) enhances data security and is trivial with Lamba Authorizers
  • Deployment automation (Terraform) facilitates quick replication across environments
  • Handling diverse data formats is crucial for comprehensive information retrieval
  • Additional tooling may be necessary for specialized tasks that LLMs are not suited for

Talk To Me

Contact Details

Need quick advice, or direction on a cloud architecture problem? Send a message and we’ll figure out a game plan. Please add as much detail as possible, and a reliable way to contact you. Thanks!

Boston Area, Massachusetts, US
@DansHardware