Case study · Summer 2025

Source-Linked YouTube RAG

A transcript-ingestion and retrieval service where every generated answer points back to its source and timestamp.

Discipline
Retrieval systems · Data engineering
Technology
Python · FastAPI · OpenAI · Milvus · REST API
TRANSCRIPTMILVUSANSWERSource 12:42 ↗

I built a FastAPI pipeline that ingests YouTube transcripts, embeds them in Milvus, and returns source-linked answers through a REST API.

TranscriptsData type
Vector searchRetrieval
Timestamp linksVerification
REST APIInterface

How the pieces connect

01Transcript
02Embeddings · Milvus
03Source-linked answer

What made the problem worth solving

Retrieval makes generated answers more grounded, but users still need a direct path back to the evidence—especially when a long video makes manual verification expensive.

Turning the problem into an engineering plan

  1. 01

    Built an ingestion pipeline for transcript segments and their source metadata.

  2. 02

    Stored embeddings in Milvus and exposed retrieval through a FastAPI service.

  3. 03

    Preserved transcript and timestamp lineage through the entire response path.

The system made auditability a first-class product behavior: the answer was useful, but the path back to the original moment in the video was equally important.

Next case studyScaffold