Search YouTube by meaning, not keywords.
InsightTube AI pulls videos, extracts their transcripts, and embeds them into a vector database. Ask in plain English and get an answer grounded in the exact videos it retrieved — with every source cited.
Built on
Attention lets each token weigh every other token, forming context-aware representations. The key insight is that queries, keys and values let the model look back at relevant earlier positions…
Sources
But what is a neural network?
3Blue1Brown
Let's build GPT: from scratch
Andrej Karpathy
Intro to LLMs & retrieval
AI Explained
/ live demo
Ask a question. Watch it retrieve.
A simulated end-to-end query. Pick a question — the engine embeds it, finds the nearest video transcripts in the vector store, then writes an answer grounded only in those sources.
retrieval trace
/ the pipeline
From a video to a grounded answer.
Every query travels the same six steps. Watch one request move through the system.
← scroll to follow the flow →
/ build your own
Drop in videos. Chat with them.
Add YouTube videos to a knowledge base — tap them on the left or paste a URL in the chat. Each one is ingested, transcribed and indexed in real time. Then ask anything, and the assistant answers using only the videos you added.
No videos yet — add some below.
tap to add
Add at least one video to your knowledge base to start asking questions.
try: youtube.com/watch?v=aircAruvnKk
/ vector space
Meaning becomes geometry.
Every video's transcript is embedded into a point in high-dimensional space. Videos that meansimilar things land close together. A query drops in and pulls its nearest neighbours — that's semantic search, visualised. Drag to orbit.
live query
/ trust & safety
Retrieval you can trust, failures you can see.
The whole point of the RAG pattern is observability. Guardrails and evaluation make sure the system gets better on purpose — not by accident.
Retrieval guardrails
Answers are constrained to retrieved context. If the store has nothing relevant, the model says so instead of guessing — wrong answers fail in a visible, debuggable way.
- Context-grounded generation
- Explicit low-confidence signals
- Source citations on every answer
Eval harness
Question-and-expected-source pairs measure whether retrieval actually got better after a model swap — so improvements are evidence-based, not vibes.
- Retrieval recall @ k
- Answer faithfulness scoring
- Regression tracking per change
API security
Every service sits behind typed contracts and standard hardening — secrets never leave the container, and requests are validated before they touch the model.
- Pydantic input validation
- Rate limiting & sane defaults
- Secrets via env, never committed
Observability
Each stage logs what it did and how long it took, so a slow or failing step is localised instantly instead of hidden inside one monolithic request.
- Per-stage latency traces
- Structured pipeline logs
- Containerised, isolated failures
/ stack
A deliberate, small stack.
Each tool does one job. They're wired together with typed contracts, so when something breaks it fails loudly and locally — not silently somewhere downstream.
from core.ai.semantic_engine import NLQueryEngine
# ask anything in plain English
engine = NLQueryEngine()
res = await engine.ask(
"best videos on vector databases?"
)
# embed → retrieve top-k → grounded answer
print(res["answer"]) # grounded in res["sources"]Run the whole pipeline locally.
Clone the repo, add your API keys, and bring up the dashboard and API in containers. Dashboard on :8501, Swagger on :8000/docs.
$ ./scripts/setup.sh && ./scripts/start.sh