Ask Mode is where Lerno started: a free-form AI tutor that behaves like chat, but is constrained to your grade, subject, and chapter and backed by retrieval from your NCERT text instead of the open web.
Below is what shipped, how it behaves, and where we are honest.
What you use it for
- Homework doubts — short questions when you are stuck on a definition, diagram, or worked example.
- Exam-style practice — ask for quick recall questions, short notes, or a slower walkthrough of a topic.
- Same-book alignment — every serious answer is built after the system looks up passages from the indexed textbook for that chapter, then explains in plain language.
Ask Mode is intentionally student-led: you choose the next question. There is no fixed lesson plan (that is what Learn Mode is for later in the roadmap).
How a session is scoped
Each Ask session is tied to one subject and one chapter (or the chapter context you opened the chat from). That matters for three reasons:
- Search stays in the right book — retrieval pulls from the correct grade and NCERT mapping, not a random mix of sources.
- Answers match what your teacher expects — vocabulary and examples line up with the chapter you are on.
- History stays organised — you can reopen a thread before a test and it still reflects that unit.
Your profile (grade, access, onboarding) still applies at the account level, but the tutor’s evidence is chapter-faithful.
What happens when you send a message (system overview)
At a high level, each user message runs through the same tutoring pipeline Lerno uses elsewhere:
- Intent — detect the kind of help you probably want (explain, notes, quiz-style prompts, solve a problem, summary) so formatting and length match the ask.
- Query preparation — your text is turned into a better search query (clearer wording for the index).
- Embeddings + hybrid search — the query is embedded; Lerno runs semantic + keyword retrieval over Qdrant chunks for that curriculum scope, then merges and ranks hits.
- Depth — a small complexity signal (quick vs standard vs detailed) keeps tiny doubts short and big questions thorough.
- Prompting — a system prompt adds subject-specific style rules, citation rules, and your learner memory (weak topics, pace, preferences) where appropriate.
- Generation — a capable model (Google Gemini in our stack) streams the reply token by token.
- Citations — when useful, the model inserts numbered references ([1], [2], …) that map back to chunk metadata (chapter, topic, pages where available).
- Persistence — the assistant turn, citations, and retrieval diagnostics are saved with the session for history and quality work.
Nothing here replaces your NCERT book — it routes the model through your book first (RAG: retrieval-augmented generation).
Capabilities shipped with Ask Mode
- Streaming replies — you read while the answer is still generating, which keeps long explanations tolerable on real networks.
- Chapter-grounded RAG — hybrid dense + lexical search, reciprocal-rank style fusion, deduplication, and capped context so prompts stay on-topic.
- Task-aware answers — different shapes for explanation vs practice vs “solve this” vs compact notes.
- Personalisation — your stored learner profile (from onboarding, chats, and practice) nudges tone and emphasis without inventing facts outside retrieved context.
- Attachments — images and PDFs can be sent with a question; vision-capable models read the page or diagram and still combine with textbook retrieval when relevant.
- Math and science rendering — responses use the same markdown + math path as the rest of the portal so equations are readable.
- Feedback and rate limits — sessions are rate-limited per user for stability; message feedback helps us improve bad turns.
Honesty and limits
- Models can be wrong — especially if retrieval misses the right chunk or your question is ambiguous. Treat citations as a pointer back to NCERT, not a guarantee.
- Not for academic dishonesty — Ask Mode is for understanding; it is not meant to ghost-write work your school expects to be wholly yours.
- Language — the stack is tuned for English and mixed English/Hinglish student queries; quality can vary on other mixes.