Descriptions:
Web Dev Cody walks through the complete architecture and implementation of semantic vector search he built for his own course platform, using PostgreSQL with the PGVector extension, OpenAI Whisper for transcription, and the text-embedding-3-small model for generating and querying embeddings. The result is a searchable index of all video transcripts that returns the most semantically relevant segments in response to natural language queries.
The pipeline runs in three stages. When a new video is uploaded, Whisper transcribes the audio (returning timestamped text stored in Postgres). A separate process chunks those transcripts into 500-token segments, sends each chunk to the OpenAI embeddings API, and stores the resulting vector alongside its segment using PGVector. At query time, the search input is embedded with the same model and compared against all stored vectors using PGVector’s built-in similarity scoring, returning a ranked list of the most relevant segments.
Cody shares the actual TypeScript/Node.js code for both the embedding creation step—using `openai.embeddings.create`—and the retrieval logic, and demonstrates a live search returning a 48% similarity match on a real query. He also covers a practical infrastructure detail: migrating to a new Postgres database on Railway was required to enable the PGVector extension, since it cannot be added to an existing instance. The video closes with a note on how understanding this architecture improves the quality of prompts given to LLMs like Opus 4.5 when building similar systems.
📺 Source: Web Dev Cody · Published January 01, 2026
🏷️ Format: Workflow Case Study
