AskMyNotes · Satyam Kumar Singh

First place, Noesis Hackathon, Vedam School of Technology (8 hours, about 50 teams)

What it does

Upload lecture notes as PDF or text, then ask questions and get answers grounded only in that material, with citations to the file and chunk and a confidence badge. Or switch to study mode and generate five MCQs or three short-answer questions from the same chunks. Voice in and out through the Web Speech API. Every question is scoped to one subject, so notes for Operating Systems can never answer a question about Networks.

From the repository

22Commits3 mine, 14%

3,437Lines of code32 source files

Counted from AskMyNotes-AlgoRise on 2026-09-15, commits 2026-02-26 to 2026-02-27. Source lines only (no vendored, generated or built files). The origin is Ved's repository; my commits are the dashboard, the initial structure and the README.

Next.js 16 App Router with three API routes (upload, ask, study), LangChain with the Google GenAI and text-splitter packages, Gemini 2.0 Flash for both embeddings and generation, a persisted ChromaDB store, and Supabase auth with middleware-protected routes.

Retrieval

Chunks of 500 characters with 50 overlap; a 0.30 similarity floor; refusal when nothing clears it.

Upload splits the text with a recursive character splitter (500 characters, 50 overlap, so a sentence cut at a chunk edge still appears whole in one of them), embeds each chunk, and stores it in Chroma with the user id, subject, file name and chunk id as metadata. Ask embeds the question, retrieves the five nearest chunks filtered by user and subject, converts Chroma’s distances to similarities, and drops anything under 0.30. If nothing survives, the route returns “Not found in your notes” without calling the model at all. Otherwise the survivors are formatted as [Source n — File, Chunk] and sent with an instruction to answer only from them or reply with that exact refusal sentence.

Confidence is the best surviving similarity, banded: High above 0.50, Medium from 0.40, Low below, and forced to Low whenever the answer contains the refusal sentence. That refusal path is the anti-hallucination layer: the system would rather say nothing than invent. The bands were chosen by feel in an eight-hour hackathon and never calibrated, which is the honest limit of the design.

Try it

The same pipeline shape, in the browser: paste notes, watch them split, ask a question, see which chunks clear the floor and what the model would be given. Ask about something the notes do not cover to hit the refusal path.

Retrieval lab500/50 chunks · top-5 · floor 0.3 · refusal path

notes.txt

Operating Systems, Unit 3: Scheduling

A process scheduler decides which ready process runs next. First-come first-served (FCFS) is non-preemptive and suffers from the convoy effect, where short processes wait behind one long process. Shortest job first (SJF) minimises average waiting time but needs the burst length in advance, which is why real systems estimate it with an exponential average of previous bursts.

Round robin (RR) gives each process a time quantum, typically 10 to 100 milliseconds. A small quantum approaches processor sharing but raises context-switch overhead; a large quantum degenerates into FCFS. The rule of thumb is that 80 percent of bursts should be shorter than the quantum.

Priority scheduling can starve low-priority processes. Aging fixes starvation by raising the priority of a process the longer it waits. Multilevel feedback queues combine these ideas: a process that uses its whole quantum moves to a lower-priority queue with a longer quantum, and one that blocks early stays high.

Deadlock is a different topic: four conditions must hold at once, mutual exclusion, hold and wait, no preemption, and circular wait. Banker's algorithm avoids deadlock by only granting a request when the resulting state is safe.

Chunks4 · 1,251 chars

chunk 0 · 415 ch Operating Systems, Unit 3: Scheduling A process scheduler decides whi…
chunk 1 · 288 ch Round robin (RR) gives each process a time quantum, typically 10 to 10…
chunk 2 · 314 ch Priority scheduling can starve low-priority processes. Aging fixes sta…
chunk 3 · 228 ch Deadlock is a different topic: four conditions must hold at once, mutu…

similaritySearchWithScoretop 5

chunk	similarity	floor
chunk 1	0.426	kept
chunk 0	0.119	dropped
chunk 2	0.069	dropped
chunk 3	0.000	dropped

Responseconfidence Medium

The model would receive the 1 surviving chunk below as context, with the instruction to answer only from them or reply exactly "Not found in your notes".

[Source 1 — File: notes.txt, Chunk: 1] Round robin (RR) gives each process a time quantum, typically 10 to 100 milliseconds. A small quantum approaches process…

In the repository the embeddings are Gemini's and the store is ChromaDB filtered by user and subject, so scores differ from the lexical coverage score used here; the mechanism (top-k, floor, refusal, citations, confidence bands) is the same. The confidence bands were never calibrated against real questions, which is the main thing I would change: see the evaluation lab for how the retrieval on this site was measured instead.

The hackathon, and who did what

Noesis, Vedam School of Technology, eight hours, about fifty teams. A rule change halfway through and a connectivity failure near submission. We were shortlisted to the top ten and then announced first. Team: Nimit Jain (landing page, frontend, Supabase auth), Ved Kumar Singh (the three API routes), and me (the dashboard, and the RAG design with Ved: ChromaDB, the chunking and overlap, the retrieval and the refusal path). The commits for the routes are Ved’s; the design of what they do was shared work.

What I took from it went straight into Ask this site: a calibrated threshold instead of a guessed one, a published question set with per-question results, and adversarial questions that must be refused. The landing page’s “98% accuracy” and “5,000 students” are placeholder marketing copy from the hackathon night, not measurements, and are not repeated here.