Megh Vyas
Backend · Infra · SRE
← All work

Retrieval engine

Multi-Modal RAG

One retrieval brain across text, images, and the open web, with a knowledge graph for the questions that need more than one hop.

Year
2025
Role
AI engineering
Built with
FastAPI, Jina embeddings, ONNX, Knowledge graph, OpenTelemetry, Kubernetes

Most retrieval systems only read text. The moment a question spans a diagram, a photo, and a web page, or needs two facts joined together, flat vector search returns something confident and wrong.

The retrieval brain

01

Encoded text, images, and web content into one shared vector space with unified cross-modal embeddings, and added a vision-language model so images are understood, not just captioned.

02

Built a knowledge graph for multi-hop reasoning, so answers that need two or three linked facts actually find them.

03

Added cross-encoder reranking on top of vector recall, so the passages that reach the model are the right ones.

04

Layered semantic caching (L1, L2, L3) and ONNX-accelerated inference to keep it fast under real load.

05

Grounded answers with source attribution and wired OpenTelemetry tracing, deployable straight onto Kubernetes.

What it answers

A production RAG engine that answers grounded, cited questions across text, images, and the open web, with the caching and tracing needed to run it at scale rather than in a notebook.

0

modalities, one index

L1-L3

semantic cache tiers

multi-hop

graph reasoning

Inside the index

Unified cross-modal embeddings (text / image / web)
Vision-language model for image understanding
Knowledge graph for multi-hop reasoning
Cross-encoder reranking over vector recall
Three-tier semantic cache (L1 / L2 / L3)
ONNX acceleration + OpenTelemetry, Kubernetes-ready