CourseAI — Smart Subtitle & Transcript Q&A

An Advanced Retrieval-Augmented Generation (RAG) system built over video course SRT and VTT subtitle files.

Ask a natural-language question about the course and get an answer that cites the exact lesson name and timestamp — so students can jump straight to the right moment in the video.

Based on the "GenAI with JavaScript 2026" course. The links in the task (Eraser design board & the course lesson) are login-protected, so this repo implements the standard Advanced RAG architecture that those lessons describe.

---

🧭 Architecture

        ┌──────────────────────────────────────────────────────────────┐
        │                        INGESTION PIPELINE                     │
        │                                                               │
  .srt  │  parse cues ─► clean/dedupe ─► chunk (w/ overlap) ─► embed    │
  .vtt  │  (parser.js)                    (chunker.js)      (llm.js)    │
        │                                        │                      │
        │                                        ▼                      │
        │                          Qdrant vector store (vectorStore.js) │
        │        payload = { lesson, source, startTs, endTs, text }     │
        └──────────────────────────────────────────────────────────────┘

        ┌──────────────────────────────────────────────────────────────┐
        │                          QUERY PIPELINE                       │
        │  question                                                     │
        │     │                                                         │
        │     ├─► Parallel Query Fan-Out ─┐                             │
        │     ├─► HyDE                    ├─► Reciprocal Rank Fusion ─►  │
        │     └─► Step-Back Prompting  ───┘        (patterns.js)        │
        │                                              │                │
        │                                              ▼                │
        │                       LLM answer w/ lesson + timestamp cites  │
        │                                   (answer.js)                 │
        └──────────────────────────────────────────────────────────────┘

Advanced RAG patterns implemented (src/patterns.js)

Pattern	What it does
Parallel Query Retrieval (Fan-Out)	LLM rewrites the question into several diverse queries, searches each, then fuses.
Reciprocal Rank Fusion (RRF)	Merges multiple ranked result lists into one robust ranking.
HyDE	Generates a hypothetical answer and searches with it for better semantic recall.
Step-Back Prompting	Asks a broader conceptual question to pull in foundational context.

---

📁 Project structure

.
├─ data/                     # put your .srt / .vtt course files here
├─ src/
│  ├─ config.js              # env + settings
│  ├─ llm.js                 # OpenAI embeddings + chat helpers
│  ├─ parser.js              # SRT/VTT -> timestamped cues
│  ├─ chunker.js             # cues -> chunks (lesson + timestamp metadata)
│  ├─ vectorStore.js         # Qdrant collection + upsert + search
│  ├─ ingest.js              # ingestion CLI
│  ├─ patterns.js            # advanced RAG retrieval strategies
│  ├─ answer.js              # grounded answer generation w/ citations
│  └─ chat.js                # query CLI / REPL
├─ docker-compose.yml        # local Qdrant
├─ .env.example
└─ package.json

---

🚀 Getting started

1. Prerequisites

• Node.js 18+
• Docker (for Qdrant) — or a hosted Qdrant Cloud URL
• An OpenAI API key

2. Install

npm install
Copy-Item .env.example .env
# then edit .env and set OPENAI_API_KEY

3. Start Qdrant

npm run qdrant:up

4. Add course subtitles

Download the SRT/VTT attachments from the course lesson and drop them into data/. File names become lesson titles, e.g. 01 - Intro to RAG.en.vtt → lesson "Intro to RAG". (Two sample files are already included so you can test immediately.)

5. Ingest

npm run ingest          # add/update
npm run ingest -- --reset   # wipe collection and re-index

6. Ask questions

# interactive
npm run chat

# one-off
node src/chat.js "How does reciprocal rank fusion work?"

# choose a strategy: all | parallel | hyde | stepback | simple
node src/chat.js --strategy hyde "What is HyDE?"

Example answer

Reciprocal Rank Fusion merges the ranked results of several search queries into a
single list, giving robustly relevant chunks (Lesson: "Chunking Subtitles", 00:00:42).

References:
- "Chunking Subtitles" @ 00:00:42–00:00:51  (02 - Chunking Subtitles.srt)
- "Introduction to RAG" @ 00:00:55–00:01:04  (01 - Introduction to RAG.en.vtt)

---

⚙️ Configuration (.env)

Variable	Default	Description
OPENAI_API_KEY	—	Required.
CHAT_MODEL	gpt-4o-mini	Chat model.
EMBEDDING_MODEL	text-embedding-3-small	Embedding model.
EMBEDDING_DIM	1536	Must match the embedding model.
QDRANT_URL	http://localhost:6333	Qdrant endpoint.
QDRANT_COLLECTION	course_subtitles	Collection name.
DATA_DIR	./data	Where subtitle files live.
CHUNK_CUES	12	Cues merged per chunk.
CHUNK_OVERLAP	2	Overlapping cues between chunks.

---

📝 Notes

• Works with auto-generated and manual subtitles; rolling-caption duplicates in auto VTT are de-duplicated automatically.
• Any OpenAI-compatible endpoint works by pointing the SDK at a different base URL.
• Swap Qdrant for another vector DB by re-implementing src/vectorStore.js.

License

MIT