You can deploy this model and start earning money today!

s0upyyy/
rag-denoiser

Name: rag-denoiser
Availability: InStock

Views

No views yet

RAG Denoising Thesis Starter (Day 1)

This repo provides a minimal, reproducible baseline RAG pipeline (E0) using SQuAD contexts as the knowledge base.

Quickstart

1) Create environment and install dependencies

bash

python -m venv .venv
# Windows:
.venv\Scripts\activate
# macOS/Linux:
# source .venv/bin/activate

pip install -U pip
pip install -r requirements.txt

2) Build corpus from SQuAD contexts

python -m src.data.squad_build_corpus --config configs/e0_clean.yaml

3) Build FAISS index

python -m src.indexing.build_faiss --config configs/e0_clean.yaml

4) Run baseline experiment (E0)

python -m src.eval.run_experiment --config configs/e0_clean.yaml --max_questions 300

Outputs will be written to outputs/.

Experiment IDs (Ablations)

E0: baseline clean
E1: baseline with noisy corpus (later)
E2: query normalization only (later)
E3: denoiser always-on (later)
E4: denoiser gated (later)
E5: dual-view retrieval (later)
E6: dual-view + reranker (later)
E7: full system (later)

Day 1 implements E0.

Day 2 (E1): Noisy Knowledge Base

E1 simulates low-level perturbations (typos / character edits) in the knowledge base text. This approximates GARAG-style corpus corruption and lets you measure how retrieval and QA degrade.

Step 1 — Generate a noisy corpus from the clean chunks

python -m src.data.make_noisy_corpus --config configs/e1_noisy_corpus.yaml

Step 2 — Build an index over the noisy corpus

python -m src.indexing.build_faiss --config configs/e1_noisy_corpus.yaml

Step 3 — Run evaluation (E1)

python -m src.eval.run_experiment --config configs/e1_noisy_corpus.yaml --max_questions 300

Outputs:

outputs/corpora/squad_chunks_noisy.jsonl
outputs/indexes/e1_noisy.faiss
outputs/indexes/e1_noisy_meta.jsonl
outputs/runs/e1_noisy_corpus/metrics.json
outputs/runs/e1_noisy_corpus/predictions.jsonl

s0upyyy/rag-denoiser