Why should you care?

Users ask natural questions. Pure keyword search misses synonyms; pure LLMs hallucinate. RAG grounded on your lessons fixes both.

See it live — copy this example

Open mongosh or MongoDB Compass, select database nosqlverse, then run the example. Change one field and run again.

// Pseudo-flow in Node
// 1) embedQuestion = await openai.embeddings(userQ)
// 2) retrieve:
const docs = await db.collection("lessons").aggregate([
  { $vectorSearch: { index: "lesson_vector_index", path: "embedding", queryVector: embedQuestion, limit: 4, numCandidates: 40 } },
  { $project: { title: 1, text: 1, score: { $meta: "vectorSearchScore" } } }
]).toArray()
// 3) prompt = `Answer using only:\n${docs.map(d => d.text).join("\n")}\nQ: ${userQ}`
// 4) llm.answer(prompt)

Run Example »

Edit the code below and click Run to see the result in Toolliyo’s live editor.

What happened?

• The question becomes a vector.
• $vectorSearch fetches relevant lesson chunks.
• The LLM answers using only those chunks.
• MongoDB is the retrieval memory; the model is the reasoner.

Practice next

1. Store lesson text chunks with embeddings.
2. Build a vector index.
3. Wire a tiny Node script: embed → vectorSearch → print docs.
4. Hybrid: $search keywords then vector re-rank.
5. Store token counts per chunk for budgeting.

Remember

RAG = retrieve from Mongo + generate with LLM. Vectors ground answers in your content. Keep chunks small and relevant.

In-app Mongo tutor

Learners ask “how does failover work?” and get answers citing the Failover lesson.

Outcome: Support load drops; answers stay on-curriculum.