Personal AI Project
Right now I am working on YouTube Chatbot, where a user can paste the url of the video and ask questions based on that. I have followed classic RAG approach. The design looks like this :
Initial design :
(query, url) → YouTube Transcript API → Translate to English (Gemini 3.5 Flash) → Chunk → Vector store (Chroma) → Similarity search → Augment context with query → LLM → Output
Upgraded design :
(query, url) → YouTube Transcript API → Chunk raw transcript → Translate to English (Gemini 3.5 Flash) asynchronously each chunk → Vector store (Chroma) → Similarity search → Augment context with query → LLM → Output
I have some intermediate steps also like if the video id is already present in vector store I will directly point to the vector store and retrieve relevant context.
There is Langsmith integration.
My main doubt here is this :
I used free gemini-3.5-flash model and it limited me to only 5 requests per minute, the problem is a particular video was 1 hr long it took approx 126 seconds to translate it using this model
I upgraded it to Tier 1 and I have changed the translation step to asynchronous, i.e, the chunks will get translated in parallel and I noticed the latency drop to 15 seconds.
I am thinking of mentioning this project in my resume, will I face any backlash because I upgraded the model?? (I think basically the model will take the same time, it is the asynchronous logic which helped in bringing the latency down to 15 sec, to make these calls happen I had to increase my Tier and get those extra calls per minute).