How ruhulamin-rakib/agrvf works
AGRVF is a notebook-based framework for verifying consumer-product claims. The documented workflow mines and prepares data, identifies products and interprets claims, retrieves local evidence, then optionally uses web evidence before credibility scoring, stance analysis, verdict selection, and explanation generation. The README documents these stages and outputs, but no source excerpts were available; subsystem-level implementation details and call ownership are therefore represented only where the documentation supports them.
Main entry points: agrvf_Layer-BCD.ipynb Layers B–D · agrvf_layer-A.ipynb Layer A · amazon_products_agrvf.parquet Amazon catalog · openbeautyfacts_agrvf.parquet Ingredient records
Components 19
Data construction
- Layer A
notebooks/agrvf_layer-A.ipynb - Claim classifier
models/claim_classifier_logreg_v1.joblib - TF-IDF vectorizer
models/tfidf_vectorizer_v1.joblib - Claim datasets
data/processed/claims - Amazon catalog
data/processed/amazon/amazon_products_agrvf.parquet - Ingredient records
data/processed/openbeautyfacts/openbeautyfacts_agrvf.parquet - Scientific corpora
data/processed
Claim understanding
- Layers B–D
notebooks/agrvf_Layer-BCD.ipynb - Product identification
notebooks/agrvf_Layer-BCD.ipynb - Claim understanding
notebooks/agrvf_Layer-BCD.ipynb - Structured claims
data/processed/layer_b/layer_b_structured_output.parquet
Evidence retrieval
- Local retrieval
notebooks/agrvf_Layer-BCD.ipynb - Local evidence
results/layer_c/layer_c_local_evidence.parquet - Web fallback
notebooks/agrvf_Layer-BCD.ipynb
Scoring and verdict
- AECS scoring
notebooks/agrvf_Layer-BCD.ipynb - Stance classification
notebooks/agrvf_Layer-BCD.ipynb - Verdict and explanation
data/processed/layer_d/layer_d_final_output.parquet
Other components
- Claim submitter
- Live web
Connections 21
- Layer A → Claim classifier: trains
- Layer A → TF-IDF vectorizer: trains
- Layer A → Claim datasets: produces
- Layer A → Amazon catalog: prepares
- Layer A → Ingredient records: prepares
- Layer A → Scientific corpora: prepares
- Claim submitter → Layers B–D: submits claim
- Layers B–D → Product identification: runs
- Layers B–D → Claim understanding: runs
- Product identification → Amazon catalog: matches against
- Layers B–D → Structured claims: writes
- Layers B–D → Local retrieval: runs
- Local retrieval → Amazon catalog: reads
- Local retrieval → Ingredient records: reads
- Local retrieval → Scientific corpora: reads
- Local retrieval → Local evidence: writes
- Layers B–D → Web fallback: runs if needed
- Web fallback → Live web: retrieves from
- Layers B–D → AECS scoring: scores evidence
- Layers B–D → Stance classification: classifies stance
- Layers B–D → Verdict and explanation: writes result