Skip to content
EpochC Start a project
· Afraz Khan

Case study: +40% query accuracy with graph + vector retrieval

A FinTech research platform (EdgeCIO) was hitting the ceiling of a vector-only retrieval approach. Complex financial questions, the kind that depend on how entities relate, returned confident but wrong answers. We rebuilt retrieval as a dual-module system and lifted query accuracy by 40%. Here is how.

The challenge

Answer complex, relationship-heavy financial queries accurately over a mix of structured and unstructured data. Pure semantic search found similar-sounding passages but missed the relationships between entities, which is exactly where the important answers lived.

What we built

A dual-module retrieval platform:

  • Neo4j knowledge graphs to model rich entity relationships (who owns what, what connects to what).
  • MongoDB vector embeddings to index unstructured content for semantic search.
  • LangChain orchestration to provide context-aware, conversational responses over both.
  • gRPC microservices with Redis caching for high-throughput, low-latency access.

Each half answered the questions the other could not: the graph for relationships and structure, the vector store for open-ended, content-based questions.

The results

  • +40% query accuracy over the legacy vector-only approach.
  • +35% answer relevance for complex financial queries.
  • Sub-200ms end-to-end retrieval latency.

What made it work

  • Hybrid beats either half alone. Relationship questions and content questions need different tools; using both covered each other’s blind spots.
  • A real ontology. We worked with data and product stakeholders to define financial ontology schemas and maintain knowledge-graph integrity, rather than dumping data into an index and hoping.
  • Latency as a first-class requirement. gRPC and Redis caching kept the system fast enough to feel conversational, not batch.

The lesson generalizes: if your RAG returns confident nonsense on relationship questions, the fix is usually not a bigger model. It is adding a knowledge graph next to your vector store.

EpochC builds hybrid retrieval systems that combine vector search and knowledge graphs. Start a project or hire on Upwork.