Open source · Runs in your browser · No paid APIs
Interview prep, grounded in your own documents.
Upload your resume, project reports and job descriptions. ResumeRAG builds a private knowledge base in your browser and prepares you for interviews with answers you can trace back to the exact passage.
The demo loads six fictional documents for “Alex Rivera”. No sign-up, nothing uploaded, and it is removed when you close the site.
You used XGBoost's scale_pos_weight parameter, because only about 7% of customers churned 1. You also tried SMOTE oversampling but dropped it: it hurt probability calibration and did not improve PR-AUC 1.
Sources
5. Modelling Approach
page 1 · Machine Learning Intern
Story 2 — Catching data leakage
How it works
Retrieval-augmented generation, step by step
An open-book exam for a language model: first find the right pages in your documents, then answer only from them.
- 01
Parse
PDF, DOCX, Markdown and text become headings, paragraphs and bullets — with page numbers.
- 02
Chunk
Structure-aware chunks that never mix two sections, each tagged with its heading path.
- 03
Embed
An open-source model turns every chunk into a 384-dimensional meaning vector — in your browser.
- 04
Hybrid search
Semantic search and BM25 keyword search, fused with reciprocal rank fusion.
- 05
Rerank
A cross-encoder re-reads the top candidates together with the question.
- 06
Generate
A language model — by default one running in this browser — answers only from the numbered passages, or refuses.
- 07
Cite & verify
Every sentence is checked against the passage it cites.
Every answer has a “How this answer was generated” panel showing each of these stages with its scores and timings.
Features
Two things to do, and tools for the rest
Not another “chat with your PDF”: ask your documents, practise answering, and reach for a specialised tool when you need one.
Ask your documents
Ask anything about your experience and get an answer built only from your own passages — each one cited by page and section, and refused outright when the evidence is not there.
Practise an interview
Questions drawn from your documents, adapting to how well you answer, with rubric feedback and a check of every claim you make against your own evidence.
Resume X-ray
Flags vague ownership, unquantified impact and expert claims, then asks the hardest legitimate questions.
Job match
Requirement by requirement. A skill only counts when your own documents prove it.
Project deep dive
Explain a project in 30 seconds or in depth, then climb a ladder of harder follow-ups.
STAR builder
Behavioural answers from real experiences, facts and suggested wording kept apart.
Consistency checker
Catches “AUC 0.91 on the resume, 0.89 in the report” before an interviewer does.
RAG lab
A retrieval playground and an evaluation suite that measure each pipeline stage.
Evaluation
Measured, not assumed
On 30 labelled questions — including deliberately unanswerable ones — each pipeline stage is scored separately. Hybrid search with cross-encoder reranking finds the needed facts 100% of the time, versus 90% for keyword search alone, and refuses every unanswerable question.
Recall@5 · MRR 0.89 with reranking · reproducible with npm run eval
Privacy
Your resume never touches our servers
Local-first by design — and honest about the one exception.
Stored on your device
Parsing, chunking, embeddings and search run in a Web Worker. Documents and vectors live in your browser's IndexedDB.
Open-weight models, in your browser
Embeddings, reranking and answer generation all run in this tab with Transformers.js — no API key, no account, no server.
Only what you choose leaves
Nothing leaves your device with the in-browser model or Ollama. Connect a hosted model and only the question plus the top passages are sent.
Walk into the interview knowing what they will ask.
Try the fictional demo in one click, then bring your own documents.