How ashokmanohar-ai/rag-evaluation-workbench-hf works
This repository provides a RAG retrieval evaluation workbench in two forms: a Gradio app and a static browser interface, plus a browser-based regression comparison page. The primary workflow ingests a corpus and labelled queries, chunks documents, ranks chunks with BM25, semantic, or hybrid retrieval, evaluates document-level ranking quality and latency, and presents downloadable provenance evidence. The Gradio implementation delegates retrieval utilities and metric/hash helpers to rag_core.py; the browser implementation performs its workflow in space-static/app.js. README capabilities are included; static page wiring beyond shown excerpts is not inferred.
Main entry points: app.py Gradio workbench · regression.js Regression page · app.js Browser workbench · rag_core.py Document chunking
Components 20
Evaluation inputs
- Corpus and labels
data - CSV input parsing
app.py - Document chunking
rag_core.py
Retrieval engines
- BM25 scoring
rag_core.py - Semantic embeddings
app.py - Transformers runtime
- Hybrid ranking
app.py - Ranked chunks
rag_core.py
Evaluation and evidence
- Document deduplication
rag_core.py - IR metrics
rag_core.py - Latency statistics
app.py - Evidence artifact
app.py
Workbench interfaces
- Gradio workbench
app.py - Browser workbench
space-static/app.js - Static interface
space-static/index.html
Regression comparison
- Regression page
space-static/regression.js - Regression interface
space-static/regression.html - Baseline comparison and gates
space-static/regression.js
Other components
- Evaluator
- Hugging Face model source
Connections 24
- Evaluator → Gradio workbench: submits inputs
app.py - Gradio workbench → CSV input parsing: parses CSV
app.py - CSV input parsing → Document chunking: supplies documents
app.py - Gradio workbench → Document chunking: chunks corpus
app.py - Gradio workbench → BM25 scoring: scores queries
app.py - Gradio workbench → Semantic embeddings: requests embeddings
app.py - Semantic embeddings → Transformers runtime: loads model
app.py - Transformers runtime → Hugging Face model source: downloads model
- Gradio workbench → Hybrid ranking: combines scores
app.py - Gradio workbench → Ranked chunks: ranks chunks
app.py - Gradio workbench → Document deduplication: deduplicates documents
app.py - Gradio workbench → IR metrics: evaluates queries
app.py - Gradio workbench → Latency statistics: aggregates latency
app.py - Gradio workbench → Evidence artifact: writes run evidence
app.py - Evaluator → Browser workbench: runs benchmark
space-static/app.js - Browser workbench → Transformers runtime: loads embeddings
space-static/app.js - Browser workbench → Hugging Face model source: fetches model
- Evaluator → Regression page: compares modes
space-static/regression.js - Regression page → Transformers runtime: loads embeddings
space-static/regression.js - Regression page → Hugging Face model source: fetches model
- Static interface → Browser workbench: hosts interface logic
- Regression interface → Regression page: hosts comparison logic
- Corpus and labels → Gradio workbench: provides example data
- Regression page → Baseline comparison and gates: compares candidate quality
space-static/regression.js