ProofRank

CPU-only hybrid ranker for the Redrob Senior AI Engineer JD: dual FAISS + dual BM25 retrieval, career-proof scoring, honeypot defenses, rule-based reasoning. No LLM, no network at rank time.

Live sandbox	https://huggingface.co/spaces/Sarthak080907/proofrank
GitHub	https://github.com/Sarthak6o1/ProofRank
Reproduce command	python rank.py --indices ./indices --out ./submission.csv

---

Reproduce the submission (Stage 3)

Pre-built indices/ ship with the repo (Git LFS). No candidates.jsonl, no build_index.py, no GPU, no network during Step 6.

Prerequisites: Git, Git LFS, Python 3.11+ on your PATH, ~2 GB free disk.
Install Python from python.org (Windows: check Add Python to PATH) before Step 5 if python is not recognized.

Clone and package install (Steps 1–5) take standard setup time depending on your network download speed.

Step 6 - the reproduce_command — runs within the hackathon compute constraint: ~10–20 s on CPU, no network, using pre-built indices/ loaded locally.

Step 1 — Enable Git LFS (once per machine)

git lfs install

Step 2 — Clone the repo

git clone https://github.com/Sarthak6o1/ProofRank.git

If Git LFS was already enabled, index files usually download during clone.

Step 3 — Enter the repo

cd ProofRank

Step 4 — Fetch index files (only if Step 6 errors on missing indices)

git lfs pull

Skip this if indices/faiss_full.index is a large file (~100+ MB), not a tiny LFS pointer stub.
LFS files: *.index, *.pkl. Normal Git files: features.parquet, candidate_ids.npy, jd_query_vec.npy.

Step 5 — Install minimal Python packages

pip install -r requirements-rank.txt

Six packages only (numpy, pandas, pyarrow, pyyaml, faiss-cpu, rank-bm25).
Do not use full requirements.txt (torch, streamlit); that is for rebuilding indices or the sandbox only.

On Windows, if python / pip fail, use py -3 and py -3 -m pip, or run scripts/setup_env.ps1 then:

.\.tools\python-embed\python.exe -m pip install -r requirements-rank.txt

If setup stops at No module named venv, that is expected — use the embed command above, then run Step 6 with .\.tools\python-embed\python.exe instead of python.

Step 6 — Generate submission CSV

This is the reproduce_command in submission_metadata.yaml. Runs offline using pre-built indices/.

python rank.py --indices ./indices --out ./submission.csv

Expected output:

Ranking with cached FAISS + BM25 + structured scorer.
Wrote 100 rows to .../submission.csv

~10–20 s on an 8-core CPU laptop — within the hackathon compute constraint (CPU-only, no network).

Step 7 — Validate CSV format (optional)

Quick local format check (stdlib only):

python India_runs_data_and_ai_challenge/validate_submission.py submission.csv

Checks header, 100 rows, ranks 1–100, monotonic scores, and CAND_####### IDs.

---

Reproduction scope

What Step 6 produces and what it depends on:

Item	Detail
Output	submission.csv: header + exactly 100 ranked rows, monotonic scores
Runtime	~10–20 s on CPU with pre-built indices/, no network
Artifacts	indices/ via Git LFS; offline index build (~85 min) is already done

Used by reproduce: rank.py, requirements-rank.txt, config/role_spec.yaml, src/*.py, indices/.

Not required for Step 6: app.py, indices_sample/, full requirements.txt, scripts/build_index.py, docs/. The Hugging Face sandbox is a separate demo path — see below.

Portal deliverables and upload checklist: SUBMISSION.md. Team metadata: submission_metadata.yaml.

---

Live sandbox (Hugging Face)

https://huggingface.co/spaces/Sarthak080907/proofrank — ranks the 50-profile sample (or your uploaded JSON ≤100) with the same engine as rank.py. Deploy notes: docs/SANDBOX_DEPLOY.md.