BridgeCom

Inspiration

During my internship, I observed a common challenge within large organizations: critical information often remains siloed within individual teams. One team might be deeply familiar with a problem, while another team may not even be aware of its existence. Connecting the dots — understanding the issue, finding the right person, and initiating the conversation — often takes considerable time and effort. This inefficiency hampers collaboration and slows down problem-solving.

This inspired the creation of BridgeCom, a system designed to break down those silos and ensure that solutions to known problems are discoverable across teams — along with the people behind those solutions.

What it does

BridgeCom is a Hybrid RAG (Retrieval-Augmented Generation) system that streamlines cross-team communication by surfacing relevant knowledge across departments. By combining Vector and Graph-based retrieval methods, it provides precise, context-aware answers from internal organizational data.

Key features:

• Retrieves answers even if the knowledge resides in another team.
• Displays the person most associated with the relevant content (with name/contact details).
• Helps avoid duplication of effort by identifying previously solved problems.
• Enables faster collaboration by suggesting the right expert to reach out to.

How we built it

• Vector Search: Used Amazon Titan embeddings and stored vectors in FAISS for fast, semantic retrieval.
• Graph Search: Modeled interdependencies using Neo4j, enhanced with retrieval from AWS Bedrock.
• Parallel RAG: Both retrieval systems run concurrently, with smart merging of context into the prompt.
• LangChain: Orchestrated prompt generation, model invocation, and response synthesis.
• Contact Linkage: Indexed metadata (person, team, contact info) to surface who solved what.

Challenges we ran into

• Initial API rate limitations with OpenAI prompted a switch to AWS Bedrock Claude, which offered better scalability.
• Crafting optimized Cypher queries to extract meaningful relationships from graph data was non-trivial.
• Synchronizing outputs from both retrieval sources (vector and graph) without losing context or consistency.
• Ensuring contact data accuracy and relevance without compromising on privacy.

Accomplishments that we're proud of

• Outperformed traditional RAG systems in ambiguous or cross-domain queries.
• Created a scalable hybrid architecture deployed entirely on AWS.
• Achieved contextual linking of problems to experts — a first step toward knowledge-aware collaboration.
• Enhanced query resolution accuracy with dynamic prompt engineering.

What we learned

• A hybrid retrieval system (vector + graph) significantly improves response relevance and context.
• Graph databases are powerful for representing relationships between people, documents, and topics.
• Embeddings are great, but people are the real connectors in problem-solving.
• Retrieval quality can be improved by enriching prompts with both content and expert metadata.

What's next for BridgeCom: Hybrid RAG using AWS Bedrock and LangChain

• Add multi-modal retrieval support (code snippets, diagrams, spreadsheets).
• Introduce streaming updates and real-time collaboration features.
• Build automated alerts when a problem similar to a past issue is detected — including who worked on it.
• Fully return to the original mission: a self-improving system that not only retrieves solutions but also connects people.

Built With

• amazon-web-services
• anthropic
• aws-bedrock
• faiss
• langchain
• llama
• neo4j