How lakshitaaaaaa/final_rag works
YouTube RAG Assistant lets users index videos, ask transcript-grounded questions with timestamp citations, and summarize a whole video. The primary flow fetches cached captions or transcribes audio when captions are unavailable, chunks the timestamped transcript, and indexes it for single-video or library-wide retrieval. Question answering checks relevance before generating a response in the question’s language; summarization instead processes the full transcript with map-reduce. The README also documents evaluation capabilities. Some links rely on README documentation or the source index; unsampled implementation details are not inferred.
Main entry points: qa_chain.py Grounded question answering · app.py Streamlit app · ingest.py Video ingestion · run_eval.py Evaluation harness
Components 16
User interface
- Streamlit app
app.py
Transcript ingestion
- Video ingestion
rag/ingest.py - Transcript fetch/cache
rag/transcript.py - Whisper fallback
rag/speech_to_text.py - Audio download/split
rag/speech_to_text.py - Transcript cache
data/transcripts - Timestamp chunking
rag/chunking.py
Retrieval and generation
- Embedding and indexing
rag/vectorstore.py - FAISS indexes
data/faiss_index - Grounded question answering
rag/qa_chain.py - Language detection
rag/language.py - Whole-video summarization
rag/summarize.py
Evaluation
- Evaluation harness
eval/run_eval.py - Evaluation metrics
eval/metrics.py
Other components
- Video viewer
- Groq API
Connections 26
- Video viewer → Streamlit app: submits video/questions
README.md - Streamlit app → Video ingestion: requests indexing
app.py - Video ingestion → Transcript fetch/cache: fetches segments
rag/ingest.py - Transcript fetch/cache → Transcript cache: reads/writes cache
README.md - Transcript fetch/cache → Whisper fallback: falls back to
rag/transcript.py - Whisper fallback → Audio download/split: downloads/splits audio
rag/speech_to_text.py - Audio download/split → Groq API: provides audio
- Whisper fallback → Groq API: requests transcription
rag/speech_to_text.py - Video ingestion → Timestamp chunking: chunks transcript
rag/ingest.py - Video ingestion → Embedding and indexing: indexes chunks
rag/ingest.py - Embedding and indexing → FAISS indexes: persists indexes
rag/vectorstore.py - Streamlit app → Embedding and indexing: loads index scope
app.py - Streamlit app → Grounded question answering: requests answers
app.py - Grounded question answering → FAISS indexes: retrieves similar chunks
README.md - Grounded question answering → Language detection: detects answer language
rag/qa_chain.py - Grounded question answering → Groq API: generates grounded answer
README.md - Streamlit app → Whole-video summarization: requests summary
app.py - Whole-video summarization → Transcript fetch/cache: reads full transcript
rag/summarize.py - Whole-video summarization → Timestamp chunking: chunks for map pass
rag/summarize.py - Whole-video summarization → Grounded question answering: uses LLM access
rag/summarize.py - Evaluation harness → Video ingestion: runs ingestion
eval/run_eval.py - Evaluation harness → Embedding and indexing: uses indexes
eval/run_eval.py - Evaluation harness → Grounded question answering: evaluates answers
eval/run_eval.py - Evaluation harness → Evaluation metrics: scores results
eval/run_eval.py - Evaluation metrics → Grounded question answering: uses LLM judge
eval/metrics.py - Evaluation metrics → Groq API: judges responses