Spaces:
Running
Running
Commit
·
c8dfbc0
0
Parent(s):
Initial commit without secrets
Browse files- .gitignore +4 -0
- Dockerfile +22 -0
- README.md +93 -0
- app/1111-main - Copy.py +41 -0
- app/222222-main - Copy.py +244 -0
- app/__init__.py +0 -0
- app/main.py +428 -0
- docker-compose.yml +12 -0
- prompts/ragbench_judge_prompt.txt +6 -0
- ragbench_eval/__init__.py +1 -0
- ragbench_eval/config.py +26 -0
- ragbench_eval/generator.py +32 -0
- ragbench_eval/judge.py +60 -0
- ragbench_eval/llm.py +47 -0
- ragbench_eval/metrics.py +77 -0
- ragbench_eval/pipeline.py +104 -0
- ragbench_eval/retriever.py +32 -0
- requirements.txt +10 -0
- scripts/run_experiment.py +35 -0
.gitignore
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
.env
|
| 2 |
+
.venv/
|
| 3 |
+
__pycache__/
|
| 4 |
+
*.pyc
|
Dockerfile
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM python:3.11-slim
|
| 2 |
+
|
| 3 |
+
WORKDIR /app
|
| 4 |
+
|
| 5 |
+
RUN apt-get update && apt-get install -y --no-install-recommends \
|
| 6 |
+
build-essential git && \
|
| 7 |
+
rm -rf /var/lib/apt/lists/*
|
| 8 |
+
|
| 9 |
+
COPY requirements.txt .
|
| 10 |
+
RUN pip install --no-cache-dir -r requirements.txt
|
| 11 |
+
|
| 12 |
+
COPY ragbench_eval ./ragbench_eval
|
| 13 |
+
COPY app ./app
|
| 14 |
+
COPY scripts ./scripts
|
| 15 |
+
COPY prompts ./prompts
|
| 16 |
+
|
| 17 |
+
ENV PYTHONUNBUFFERED=1
|
| 18 |
+
|
| 19 |
+
# Hugging Face Spaces expect 7860
|
| 20 |
+
EXPOSE 7860
|
| 21 |
+
|
| 22 |
+
CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "7860"]
|
README.md
ADDED
|
@@ -0,0 +1,93 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
title: ragbench-rag-eval
|
| 3 |
+
emoji: "📊"
|
| 4 |
+
colorFrom: blue
|
| 5 |
+
colorTo: indigo
|
| 6 |
+
sdk: docker
|
| 7 |
+
pinned: false
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# RAGBench RAG Evaluation Project
|
| 11 |
+
|
| 12 |
+
This project evaluates a RAG system on the RAGBench dataset across 5 domains:
|
| 13 |
+
Biomedical, General Knowledge, Legal, Customer Support, and Finance.
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
# RAGBench RAG Evaluation Project
|
| 17 |
+
|
| 18 |
+
This project evaluates a RAG system on the RAGBench dataset across 5 domains:
|
| 19 |
+
Biomedical, General Knowledge, Legal, Customer Support, and Finance.
|
| 20 |
+
|
| 21 |
+
## 1. Setup (local, no Docker)
|
| 22 |
+
|
| 23 |
+
```bash
|
| 24 |
+
python -m venv .venv
|
| 25 |
+
source .venv/bin/activate # Windows: .venv\\Scripts\\activate
|
| 26 |
+
pip install --upgrade pip
|
| 27 |
+
pip install -r requirements.txt
|
| 28 |
+
```
|
| 29 |
+
|
| 30 |
+
Copy `.env.example` to `.env` and fill in:
|
| 31 |
+
|
| 32 |
+
- HF_TOKEN (if using Hugging Face models)
|
| 33 |
+
- GROQ_API_KEY (if using Groq)
|
| 34 |
+
- RAGBENCH_LLM_PROVIDER = groq or hf
|
| 35 |
+
- RAGBENCH_GEN_MODEL
|
| 36 |
+
- RAGBENCH_JUDGE_MODEL
|
| 37 |
+
|
| 38 |
+
Also open `prompts/ragbench_judge_prompt.txt` and paste the official JSON
|
| 39 |
+
annotation prompt from the RAGBench paper (Appendix 9.4), with placeholders:
|
| 40 |
+
`{documents}`, `{question}`, `{answer}`.
|
| 41 |
+
|
| 42 |
+
### Run an experiment from CLI
|
| 43 |
+
|
| 44 |
+
```bash
|
| 45 |
+
python -m scripts.run_experiment --domain biomedical --k 3 --max_examples 10
|
| 46 |
+
```
|
| 47 |
+
|
| 48 |
+
## 2. Run FastAPI locally (no Docker)
|
| 49 |
+
|
| 50 |
+
```bash
|
| 51 |
+
uvicorn app.main:app --host 0.0.0.0 --port 7860
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
Then open:
|
| 55 |
+
|
| 56 |
+
- `http://localhost:7860/health`
|
| 57 |
+
- `http://localhost:7860/docs` (Swagger UI)
|
| 58 |
+
- POST `/run_domain` with JSON:
|
| 59 |
+
|
| 60 |
+
```json
|
| 61 |
+
{
|
| 62 |
+
"domain": "biomedical",
|
| 63 |
+
"k": 3,
|
| 64 |
+
"max_examples": 10,
|
| 65 |
+
"split": "test"
|
| 66 |
+
}
|
| 67 |
+
```
|
| 68 |
+
|
| 69 |
+
## 3. Run with Docker (local laptop)
|
| 70 |
+
|
| 71 |
+
Build and run:
|
| 72 |
+
|
| 73 |
+
```bash
|
| 74 |
+
docker compose build
|
| 75 |
+
docker compose up
|
| 76 |
+
```
|
| 77 |
+
|
| 78 |
+
The API will be available at `http://localhost:8000`.
|
| 79 |
+
|
| 80 |
+
## 4. Deploy to Hugging Face Space (Docker)
|
| 81 |
+
|
| 82 |
+
1. Create a new Space with SDK = Docker.
|
| 83 |
+
2. Push this repo to the Space Git URL.
|
| 84 |
+
3. On the Space settings, add variables/secrets:
|
| 85 |
+
|
| 86 |
+
- HF_TOKEN
|
| 87 |
+
- GROQ_API_KEY
|
| 88 |
+
- RAGBENCH_LLM_PROVIDER
|
| 89 |
+
- RAGBENCH_GEN_MODEL
|
| 90 |
+
- RAGBENCH_JUDGE_MODEL
|
| 91 |
+
|
| 92 |
+
4. Once the Space builds successfully, open `/docs` on the Space URL to run
|
| 93 |
+
`/run_domain` for each domain via Swagger UI.
|
app/1111-main - Copy.py
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from fastapi import FastAPI
|
| 2 |
+
from pydantic import BaseModel
|
| 3 |
+
|
| 4 |
+
from ragbench_eval.pipeline import RagBenchExperiment
|
| 5 |
+
|
| 6 |
+
app = FastAPI(title="RAGBench RAG Evaluation API")
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class RunRequest(BaseModel):
|
| 10 |
+
domain: str
|
| 11 |
+
k: int = 3
|
| 12 |
+
max_examples: int = 20
|
| 13 |
+
split: str = "test"
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
@app.post("/run_domain")
|
| 17 |
+
def run_domain(req: RunRequest):
|
| 18 |
+
exp = RagBenchExperiment(
|
| 19 |
+
k=req.k,
|
| 20 |
+
max_examples=req.max_examples,
|
| 21 |
+
split=req.split,
|
| 22 |
+
)
|
| 23 |
+
result = exp.run_domain(req.domain)
|
| 24 |
+
return result
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
@app.get("/health")
|
| 28 |
+
def health():
|
| 29 |
+
return {"status": "ok"}
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
@app.get("/")
|
| 33 |
+
def root():
|
| 34 |
+
return {
|
| 35 |
+
"message": "RAGBench RAG Evaluation API is running.",
|
| 36 |
+
"endpoints": {
|
| 37 |
+
"health": "/health",
|
| 38 |
+
"docs": "/docs",
|
| 39 |
+
"run_domain": "/run_domain (POST)",
|
| 40 |
+
},
|
| 41 |
+
}
|
app/222222-main - Copy.py
ADDED
|
@@ -0,0 +1,244 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from fastapi import FastAPI
|
| 2 |
+
from fastapi.responses import HTMLResponse
|
| 3 |
+
from pydantic import BaseModel
|
| 4 |
+
|
| 5 |
+
from ragbench_eval.pipeline import RagBenchExperiment
|
| 6 |
+
|
| 7 |
+
app = FastAPI(title="RAGBench RAG Evaluation API")
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class RunRequest(BaseModel):
|
| 11 |
+
domain: str
|
| 12 |
+
k: int = 3
|
| 13 |
+
max_examples: int = 20
|
| 14 |
+
split: str = "test"
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
@app.post("/run_domain")
|
| 18 |
+
def run_domain(req: RunRequest):
|
| 19 |
+
exp = RagBenchExperiment(
|
| 20 |
+
k=req.k,
|
| 21 |
+
max_examples=req.max_examples,
|
| 22 |
+
split=req.split,
|
| 23 |
+
)
|
| 24 |
+
result = exp.run_domain(req.domain)
|
| 25 |
+
return result
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
@app.get("/health")
|
| 29 |
+
def health():
|
| 30 |
+
return {"status": "ok"}
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
@app.get("/")
|
| 34 |
+
def root():
|
| 35 |
+
return {
|
| 36 |
+
"message": "RAGBench RAG Evaluation API is running.",
|
| 37 |
+
"endpoints": {
|
| 38 |
+
"health": "/health",
|
| 39 |
+
"docs": "/docs",
|
| 40 |
+
"ui": "/ui",
|
| 41 |
+
"run_domain": "/run_domain (POST)",
|
| 42 |
+
},
|
| 43 |
+
}
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
# ------------- NEW: simple frontend at /ui -----------------
|
| 47 |
+
|
| 48 |
+
@app.get("/ui", response_class=HTMLResponse)
|
| 49 |
+
def ui():
|
| 50 |
+
html = """
|
| 51 |
+
<!DOCTYPE html>
|
| 52 |
+
<html lang="en">
|
| 53 |
+
<head>
|
| 54 |
+
<meta charset="UTF-8" />
|
| 55 |
+
<title>RAGBench RAG Evaluation UI</title>
|
| 56 |
+
<meta name="viewport" content="width=device-width, initial-scale=1" />
|
| 57 |
+
<style>
|
| 58 |
+
body {
|
| 59 |
+
font-family: system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
|
| 60 |
+
margin: 0;
|
| 61 |
+
padding: 0;
|
| 62 |
+
background: #f5f7fa;
|
| 63 |
+
color: #111827;
|
| 64 |
+
}
|
| 65 |
+
.wrapper {
|
| 66 |
+
max-width: 960px;
|
| 67 |
+
margin: 2rem auto;
|
| 68 |
+
padding: 1.5rem;
|
| 69 |
+
background: #ffffff;
|
| 70 |
+
border-radius: 0.75rem;
|
| 71 |
+
box-shadow: 0 10px 25px rgba(0, 0, 0, 0.06);
|
| 72 |
+
}
|
| 73 |
+
h1 {
|
| 74 |
+
margin-top: 0;
|
| 75 |
+
font-size: 1.6rem;
|
| 76 |
+
}
|
| 77 |
+
.row {
|
| 78 |
+
display: flex;
|
| 79 |
+
flex-wrap: wrap;
|
| 80 |
+
gap: 1rem;
|
| 81 |
+
margin-bottom: 1rem;
|
| 82 |
+
}
|
| 83 |
+
.field {
|
| 84 |
+
flex: 1 1 180px;
|
| 85 |
+
min-width: 160px;
|
| 86 |
+
}
|
| 87 |
+
label {
|
| 88 |
+
display: block;
|
| 89 |
+
font-size: 0.85rem;
|
| 90 |
+
font-weight: 600;
|
| 91 |
+
margin-bottom: 0.25rem;
|
| 92 |
+
}
|
| 93 |
+
select, input {
|
| 94 |
+
width: 100%;
|
| 95 |
+
padding: 0.45rem 0.55rem;
|
| 96 |
+
border-radius: 0.375rem;
|
| 97 |
+
border: 1px solid #d1d5db;
|
| 98 |
+
font-size: 0.9rem;
|
| 99 |
+
box-sizing: border-box;
|
| 100 |
+
}
|
| 101 |
+
button {
|
| 102 |
+
padding: 0.55rem 1.2rem;
|
| 103 |
+
border-radius: 999px;
|
| 104 |
+
border: none;
|
| 105 |
+
background: #2563eb;
|
| 106 |
+
color: #ffffff;
|
| 107 |
+
font-weight: 600;
|
| 108 |
+
font-size: 0.95rem;
|
| 109 |
+
cursor: pointer;
|
| 110 |
+
}
|
| 111 |
+
button:disabled {
|
| 112 |
+
opacity: 0.6;
|
| 113 |
+
cursor: default;
|
| 114 |
+
}
|
| 115 |
+
.actions {
|
| 116 |
+
margin-top: 0.5rem;
|
| 117 |
+
margin-bottom: 1rem;
|
| 118 |
+
}
|
| 119 |
+
.status {
|
| 120 |
+
font-size: 0.85rem;
|
| 121 |
+
margin-bottom: 0.5rem;
|
| 122 |
+
color: #4b5563;
|
| 123 |
+
}
|
| 124 |
+
pre {
|
| 125 |
+
background: #0b1020;
|
| 126 |
+
color: #e5e7eb;
|
| 127 |
+
padding: 1rem;
|
| 128 |
+
border-radius: 0.75rem;
|
| 129 |
+
overflow: auto;
|
| 130 |
+
max-height: 480px;
|
| 131 |
+
font-size: 0.8rem;
|
| 132 |
+
}
|
| 133 |
+
code {
|
| 134 |
+
font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", "Courier New", monospace;
|
| 135 |
+
}
|
| 136 |
+
@media (max-width: 640px) {
|
| 137 |
+
.wrapper {
|
| 138 |
+
margin: 0.5rem;
|
| 139 |
+
border-radius: 0.5rem;
|
| 140 |
+
}
|
| 141 |
+
}
|
| 142 |
+
</style>
|
| 143 |
+
</head>
|
| 144 |
+
<body>
|
| 145 |
+
<div class="wrapper">
|
| 146 |
+
<h1>RAGBench RAG Evaluation</h1>
|
| 147 |
+
<p style="font-size:0.9rem; color:#4b5563;">
|
| 148 |
+
Use this UI to call <code>POST /run_domain</code> and inspect the metrics
|
| 149 |
+
for a given domain. The backend uses the RAGBench dataset and your configured LLMs.
|
| 150 |
+
</p>
|
| 151 |
+
|
| 152 |
+
<div class="row">
|
| 153 |
+
<div class="field">
|
| 154 |
+
<label for="domain">Domain</label>
|
| 155 |
+
<select id="domain">
|
| 156 |
+
<option value="biomedical">Biomedical</option>
|
| 157 |
+
<option value="general_knowledge">General Knowledge</option>
|
| 158 |
+
<option value="legal">Legal</option>
|
| 159 |
+
<option value="customer_support">Customer Support</option>
|
| 160 |
+
<option value="finance">Finance</option>
|
| 161 |
+
</select>
|
| 162 |
+
</div>
|
| 163 |
+
|
| 164 |
+
<div class="field">
|
| 165 |
+
<label for="k">Top-k documents</label>
|
| 166 |
+
<input id="k" type="number" value="3" min="1" />
|
| 167 |
+
</div>
|
| 168 |
+
|
| 169 |
+
<div class="field">
|
| 170 |
+
<label for="max_examples">Max examples</label>
|
| 171 |
+
<input id="max_examples" type="number" value="5" min="1" />
|
| 172 |
+
</div>
|
| 173 |
+
|
| 174 |
+
<div class="field">
|
| 175 |
+
<label for="split">Dataset split</label>
|
| 176 |
+
<input id="split" type="text" value="test" />
|
| 177 |
+
</div>
|
| 178 |
+
</div>
|
| 179 |
+
|
| 180 |
+
<div class="actions">
|
| 181 |
+
<button id="runBtn" onclick="runDomain()">Run Domain Evaluation</button>
|
| 182 |
+
</div>
|
| 183 |
+
|
| 184 |
+
<div class="status" id="status"></div>
|
| 185 |
+
|
| 186 |
+
<pre><code id="output">{}</code></pre>
|
| 187 |
+
</div>
|
| 188 |
+
|
| 189 |
+
<script>
|
| 190 |
+
async function runDomain() {
|
| 191 |
+
const domainEl = document.getElementById("domain");
|
| 192 |
+
const kEl = document.getElementById("k");
|
| 193 |
+
const maxExamplesEl = document.getElementById("max_examples");
|
| 194 |
+
const splitEl = document.getElementById("split");
|
| 195 |
+
const statusEl = document.getElementById("status");
|
| 196 |
+
const outputEl = document.getElementById("output");
|
| 197 |
+
const btn = document.getElementById("runBtn");
|
| 198 |
+
|
| 199 |
+
const domain = domainEl.value;
|
| 200 |
+
const k = parseInt(kEl.value || "3", 10);
|
| 201 |
+
const maxExamples = parseInt(maxExamplesEl.value || "5", 10);
|
| 202 |
+
const split = splitEl.value || "test";
|
| 203 |
+
|
| 204 |
+
const payload = {
|
| 205 |
+
domain: domain,
|
| 206 |
+
k: k,
|
| 207 |
+
max_examples: maxExamples,
|
| 208 |
+
split: split
|
| 209 |
+
};
|
| 210 |
+
|
| 211 |
+
statusEl.textContent = "Running evaluation...";
|
| 212 |
+
btn.disabled = true;
|
| 213 |
+
outputEl.textContent = "{}";
|
| 214 |
+
|
| 215 |
+
try {
|
| 216 |
+
const res = await fetch("/run_domain", {
|
| 217 |
+
method: "POST",
|
| 218 |
+
headers: {
|
| 219 |
+
"Content-Type": "application/json"
|
| 220 |
+
},
|
| 221 |
+
body: JSON.stringify(payload)
|
| 222 |
+
});
|
| 223 |
+
|
| 224 |
+
const data = await res.json();
|
| 225 |
+
|
| 226 |
+
if (!res.ok) {
|
| 227 |
+
statusEl.textContent = "Error " + res.status;
|
| 228 |
+
} else {
|
| 229 |
+
statusEl.textContent = "Done.";
|
| 230 |
+
}
|
| 231 |
+
|
| 232 |
+
outputEl.textContent = JSON.stringify(data, null, 2);
|
| 233 |
+
} catch (err) {
|
| 234 |
+
statusEl.textContent = "Request failed: " + err;
|
| 235 |
+
outputEl.textContent = "{}";
|
| 236 |
+
} finally {
|
| 237 |
+
btn.disabled = false;
|
| 238 |
+
}
|
| 239 |
+
}
|
| 240 |
+
</script>
|
| 241 |
+
</body>
|
| 242 |
+
</html>
|
| 243 |
+
"""
|
| 244 |
+
return HTMLResponse(content=html)
|
app/__init__.py
ADDED
|
File without changes
|
app/main.py
ADDED
|
@@ -0,0 +1,428 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import List, Tuple
|
| 2 |
+
|
| 3 |
+
from fastapi import FastAPI
|
| 4 |
+
from fastapi.responses import HTMLResponse
|
| 5 |
+
from pydantic import BaseModel
|
| 6 |
+
from datasets import load_dataset
|
| 7 |
+
|
| 8 |
+
from ragbench_eval.pipeline import RagBenchExperiment
|
| 9 |
+
from ragbench_eval.retriever import ExampleRetriever
|
| 10 |
+
from ragbench_eval.generator import RAGGenerator
|
| 11 |
+
from ragbench_eval.judge import RAGJudge
|
| 12 |
+
from ragbench_eval.metrics import trace_from_attributes
|
| 13 |
+
from ragbench_eval.config import RAGBENCH_DATASET
|
| 14 |
+
|
| 15 |
+
app = FastAPI(title="RAGBench RAG Evaluation API")
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class RunRequest(BaseModel):
|
| 19 |
+
domain: str
|
| 20 |
+
k: int = 3
|
| 21 |
+
max_examples: int = 20
|
| 22 |
+
split: str = "test"
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class QAExampleRequest(BaseModel):
|
| 26 |
+
subset: str # e.g. "covidqa", "pubmedqa", "finqa"
|
| 27 |
+
index: int = 0 # which example in that subset
|
| 28 |
+
k: int = 3 # top-k docs
|
| 29 |
+
split: str = "test" # usually "test"
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
@app.post("/run_domain")
|
| 33 |
+
def run_domain(req: RunRequest):
|
| 34 |
+
exp = RagBenchExperiment(
|
| 35 |
+
k=req.k,
|
| 36 |
+
max_examples=req.max_examples,
|
| 37 |
+
split=req.split,
|
| 38 |
+
)
|
| 39 |
+
result = exp.run_domain(req.domain)
|
| 40 |
+
return result
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
@app.post("/qa_example")
|
| 44 |
+
def qa_example(req: QAExampleRequest):
|
| 45 |
+
"""
|
| 46 |
+
Run RAG on a single RAGBench example and return:
|
| 47 |
+
- question
|
| 48 |
+
- generated answer
|
| 49 |
+
- retrieved docs with sentence keys
|
| 50 |
+
- judge attributes
|
| 51 |
+
- predicted TRACe metrics
|
| 52 |
+
- ground-truth scores from dataset
|
| 53 |
+
"""
|
| 54 |
+
ds = load_dataset(RAGBENCH_DATASET, req.subset, split=req.split)
|
| 55 |
+
|
| 56 |
+
if req.index < 0 or req.index >= len(ds):
|
| 57 |
+
return {"error": f"index {req.index} out of range (0..{len(ds)-1})"}
|
| 58 |
+
|
| 59 |
+
row = ds[req.index]
|
| 60 |
+
|
| 61 |
+
docs_sentences_full: List[List[Tuple[str, str]]] = []
|
| 62 |
+
for doc in row["documents_sentences"]:
|
| 63 |
+
docs_sentences_full.append([(k, s) for k, s in doc])
|
| 64 |
+
|
| 65 |
+
question = row["question"]
|
| 66 |
+
|
| 67 |
+
retriever = ExampleRetriever()
|
| 68 |
+
doc_indices = retriever.rank_docs(question, docs_sentences_full, k=req.k)
|
| 69 |
+
selected_docs = [docs_sentences_full[j] for j in doc_indices]
|
| 70 |
+
|
| 71 |
+
generator = RAGGenerator()
|
| 72 |
+
answer = generator.generate(question, selected_docs)
|
| 73 |
+
|
| 74 |
+
judge = RAGJudge()
|
| 75 |
+
attrs = judge.annotate(question, answer, selected_docs)
|
| 76 |
+
|
| 77 |
+
pred_metrics = trace_from_attributes(attrs, selected_docs)
|
| 78 |
+
|
| 79 |
+
docs_view = []
|
| 80 |
+
for doc_i, doc in enumerate(selected_docs):
|
| 81 |
+
docs_view.append({
|
| 82 |
+
"doc_index": doc_indices[doc_i],
|
| 83 |
+
"sentences": [{"key": k, "text": s} for k, s in doc],
|
| 84 |
+
})
|
| 85 |
+
|
| 86 |
+
return {
|
| 87 |
+
"subset": req.subset,
|
| 88 |
+
"index": req.index,
|
| 89 |
+
"question": question,
|
| 90 |
+
"answer": answer,
|
| 91 |
+
"retrieved_docs": docs_view,
|
| 92 |
+
"judge_attributes": attrs,
|
| 93 |
+
"predicted_trace_metrics": pred_metrics,
|
| 94 |
+
"ground_truth": {
|
| 95 |
+
"relevance_score": row.get("relevance_score"),
|
| 96 |
+
"utilization_score": row.get("utilization_score"),
|
| 97 |
+
"completeness_score": row.get("completeness_score"),
|
| 98 |
+
"adherence_score": row.get("adherence_score"),
|
| 99 |
+
},
|
| 100 |
+
}
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
@app.get("/health")
|
| 104 |
+
def health():
|
| 105 |
+
return {"status": "ok"}
|
| 106 |
+
|
| 107 |
+
|
| 108 |
+
@app.get("/")
|
| 109 |
+
def root():
|
| 110 |
+
return {
|
| 111 |
+
"message": "RAGBench RAG Evaluation API is running.",
|
| 112 |
+
"endpoints": {
|
| 113 |
+
"health": "/health",
|
| 114 |
+
"docs": "/docs",
|
| 115 |
+
"ui": "/ui",
|
| 116 |
+
"run_domain": "/run_domain (POST)",
|
| 117 |
+
"qa_example": "/qa_example (POST)",
|
| 118 |
+
},
|
| 119 |
+
}
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
@app.get("/ui", response_class=HTMLResponse)
|
| 123 |
+
def ui():
|
| 124 |
+
html = """
|
| 125 |
+
<!DOCTYPE html>
|
| 126 |
+
<html lang="en">
|
| 127 |
+
<head>
|
| 128 |
+
<meta charset="UTF-8" />
|
| 129 |
+
<title>RAGBench RAG Evaluation UI</title>
|
| 130 |
+
<meta name="viewport" content="width=device-width, initial-scale=1" />
|
| 131 |
+
<style>
|
| 132 |
+
body {
|
| 133 |
+
font-family: system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif;
|
| 134 |
+
margin: 0;
|
| 135 |
+
padding: 0;
|
| 136 |
+
background: #f3f4f6;
|
| 137 |
+
color: #111827;
|
| 138 |
+
}
|
| 139 |
+
.wrapper {
|
| 140 |
+
max-width: 1080px;
|
| 141 |
+
margin: 2rem auto;
|
| 142 |
+
padding: 1.5rem;
|
| 143 |
+
}
|
| 144 |
+
.card {
|
| 145 |
+
background: #ffffff;
|
| 146 |
+
border-radius: 0.75rem;
|
| 147 |
+
box-shadow: 0 10px 25px rgba(0, 0, 0, 0.06);
|
| 148 |
+
padding: 1.25rem 1.5rem;
|
| 149 |
+
margin-bottom: 1.5rem;
|
| 150 |
+
}
|
| 151 |
+
h1 {
|
| 152 |
+
margin-top: 0;
|
| 153 |
+
font-size: 1.6rem;
|
| 154 |
+
}
|
| 155 |
+
h2 {
|
| 156 |
+
margin-top: 0;
|
| 157 |
+
font-size: 1.2rem;
|
| 158 |
+
}
|
| 159 |
+
p {
|
| 160 |
+
font-size: 0.9rem;
|
| 161 |
+
color: #4b5563;
|
| 162 |
+
}
|
| 163 |
+
.row {
|
| 164 |
+
display: flex;
|
| 165 |
+
flex-wrap: wrap;
|
| 166 |
+
gap: 1rem;
|
| 167 |
+
margin-bottom: 1rem;
|
| 168 |
+
}
|
| 169 |
+
.field {
|
| 170 |
+
flex: 1 1 180px;
|
| 171 |
+
min-width: 160px;
|
| 172 |
+
}
|
| 173 |
+
label {
|
| 174 |
+
display: block;
|
| 175 |
+
font-size: 0.85rem;
|
| 176 |
+
font-weight: 600;
|
| 177 |
+
margin-bottom: 0.25rem;
|
| 178 |
+
}
|
| 179 |
+
select, input {
|
| 180 |
+
width: 100%;
|
| 181 |
+
padding: 0.45rem 0.55rem;
|
| 182 |
+
border-radius: 0.375rem;
|
| 183 |
+
border: 1px solid #d1d5db;
|
| 184 |
+
font-size: 0.9rem;
|
| 185 |
+
box-sizing: border-box;
|
| 186 |
+
}
|
| 187 |
+
button {
|
| 188 |
+
padding: 0.55rem 1.2rem;
|
| 189 |
+
border-radius: 999px;
|
| 190 |
+
border: none;
|
| 191 |
+
background: #2563eb;
|
| 192 |
+
color: #ffffff;
|
| 193 |
+
font-weight: 600;
|
| 194 |
+
font-size: 0.95rem;
|
| 195 |
+
cursor: pointer;
|
| 196 |
+
}
|
| 197 |
+
button:disabled {
|
| 198 |
+
opacity: 0.6;
|
| 199 |
+
cursor: default;
|
| 200 |
+
}
|
| 201 |
+
.actions {
|
| 202 |
+
margin-top: 0.5rem;
|
| 203 |
+
margin-bottom: 0.75rem;
|
| 204 |
+
}
|
| 205 |
+
.status {
|
| 206 |
+
font-size: 0.85rem;
|
| 207 |
+
margin-bottom: 0.5rem;
|
| 208 |
+
color: #4b5563;
|
| 209 |
+
}
|
| 210 |
+
pre {
|
| 211 |
+
background: #0b1020;
|
| 212 |
+
color: #e5e7eb;
|
| 213 |
+
padding: 1rem;
|
| 214 |
+
border-radius: 0.75rem;
|
| 215 |
+
overflow: auto;
|
| 216 |
+
max-height: 420px;
|
| 217 |
+
font-size: 0.8rem;
|
| 218 |
+
}
|
| 219 |
+
code {
|
| 220 |
+
font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", "Courier New", monospace;
|
| 221 |
+
}
|
| 222 |
+
@media (max-width: 640px) {
|
| 223 |
+
.wrapper {
|
| 224 |
+
margin: 0.5rem;
|
| 225 |
+
padding: 0.75rem;
|
| 226 |
+
}
|
| 227 |
+
.card {
|
| 228 |
+
padding: 0.9rem 1rem;
|
| 229 |
+
}
|
| 230 |
+
}
|
| 231 |
+
</style>
|
| 232 |
+
</head>
|
| 233 |
+
<body>
|
| 234 |
+
<div class="wrapper">
|
| 235 |
+
<div class="card">
|
| 236 |
+
<h1>RAGBench RAG Evaluation</h1>
|
| 237 |
+
<p>
|
| 238 |
+
This UI lets you:
|
| 239 |
+
(1) run domain-level evaluation on RAGBench, and
|
| 240 |
+
(2) inspect a single example (question, retrieved docs, answer, and metrics).
|
| 241 |
+
</p>
|
| 242 |
+
</div>
|
| 243 |
+
|
| 244 |
+
<!-- Domain evaluation card -->
|
| 245 |
+
<div class="card">
|
| 246 |
+
<h2>1. Domain Evaluation (POST /run_domain)</h2>
|
| 247 |
+
<p>
|
| 248 |
+
Evaluate all subsets in a domain using the configured LLM and retriever.
|
| 249 |
+
</p>
|
| 250 |
+
|
| 251 |
+
<div class="row">
|
| 252 |
+
<div class="field">
|
| 253 |
+
<label for="domain">Domain</label>
|
| 254 |
+
<select id="domain">
|
| 255 |
+
<option value="biomedical">Biomedical</option>
|
| 256 |
+
<option value="general_knowledge">General Knowledge</option>
|
| 257 |
+
<option value="legal">Legal</option>
|
| 258 |
+
<option value="customer_support">Customer Support</option>
|
| 259 |
+
<option value="finance">Finance</option>
|
| 260 |
+
</select>
|
| 261 |
+
</div>
|
| 262 |
+
|
| 263 |
+
<div class="field">
|
| 264 |
+
<label for="k">Top-k documents</label>
|
| 265 |
+
<input id="k" type="number" value="3" min="1" />
|
| 266 |
+
</div>
|
| 267 |
+
|
| 268 |
+
<div class="field">
|
| 269 |
+
<label for="max_examples">Max examples</label>
|
| 270 |
+
<input id="max_examples" type="number" value="5" min="1" />
|
| 271 |
+
</div>
|
| 272 |
+
|
| 273 |
+
<div class="field">
|
| 274 |
+
<label for="split">Dataset split</label>
|
| 275 |
+
<input id="split" type="text" value="test" />
|
| 276 |
+
</div>
|
| 277 |
+
</div>
|
| 278 |
+
|
| 279 |
+
<div class="actions">
|
| 280 |
+
<button id="runBtn" onclick="runDomain()">Run Domain Evaluation</button>
|
| 281 |
+
</div>
|
| 282 |
+
|
| 283 |
+
<div class="status" id="status"></div>
|
| 284 |
+
<pre><code id="output">{}</code></pre>
|
| 285 |
+
</div>
|
| 286 |
+
|
| 287 |
+
<!-- Single example viewer card -->
|
| 288 |
+
<div class="card">
|
| 289 |
+
<h2>2. Single Example Viewer (POST /qa_example)</h2>
|
| 290 |
+
<p>
|
| 291 |
+
Inspect one RAGBench example: question, retrieved documents, answer,
|
| 292 |
+
judge attributes, and TRACe metrics.
|
| 293 |
+
</p>
|
| 294 |
+
|
| 295 |
+
<div class="row">
|
| 296 |
+
<div class="field">
|
| 297 |
+
<label for="subset">Subset</label>
|
| 298 |
+
<input list="subset-list" id="subset" value="covidqa" />
|
| 299 |
+
<datalist id="subset-list">
|
| 300 |
+
<option value="pubmedqa">
|
| 301 |
+
<option value="covidqa">
|
| 302 |
+
<option value="hotpotqa">
|
| 303 |
+
<option value="msmarco">
|
| 304 |
+
<option value="hagrid">
|
| 305 |
+
<option value="expertqa">
|
| 306 |
+
<option value="cuad">
|
| 307 |
+
<option value="delucionqa">
|
| 308 |
+
<option value="emanual">
|
| 309 |
+
<option value="techqa">
|
| 310 |
+
<option value="finqa">
|
| 311 |
+
<option value="tatqa">
|
| 312 |
+
</datalist>
|
| 313 |
+
</div>
|
| 314 |
+
|
| 315 |
+
<div class="field">
|
| 316 |
+
<label for="example_index">Example index</label>
|
| 317 |
+
<input id="example_index" type="number" value="0" min="0" />
|
| 318 |
+
</div>
|
| 319 |
+
|
| 320 |
+
<div class="field">
|
| 321 |
+
<label for="k_example">Top-k documents</label>
|
| 322 |
+
<input id="k_example" type="number" value="3" min="1" />
|
| 323 |
+
</div>
|
| 324 |
+
|
| 325 |
+
<div class="field">
|
| 326 |
+
<label for="split_example">Dataset split</label>
|
| 327 |
+
<input id="split_example" type="text" value="test" />
|
| 328 |
+
</div>
|
| 329 |
+
</div>
|
| 330 |
+
|
| 331 |
+
<div class="actions">
|
| 332 |
+
<button id="qaBtn" onclick="runExample()">Run Single Example</button>
|
| 333 |
+
</div>
|
| 334 |
+
|
| 335 |
+
<div class="status" id="qa_status"></div>
|
| 336 |
+
<pre><code id="qa_output">{}</code></pre>
|
| 337 |
+
</div>
|
| 338 |
+
</div>
|
| 339 |
+
|
| 340 |
+
<script>
|
| 341 |
+
async function runDomain() {
|
| 342 |
+
const domainEl = document.getElementById("domain");
|
| 343 |
+
const kEl = document.getElementById("k");
|
| 344 |
+
const maxExamplesEl = document.getElementById("max_examples");
|
| 345 |
+
const splitEl = document.getElementById("split");
|
| 346 |
+
const statusEl = document.getElementById("status");
|
| 347 |
+
const outputEl = document.getElementById("output");
|
| 348 |
+
const btn = document.getElementById("runBtn");
|
| 349 |
+
|
| 350 |
+
const payload = {
|
| 351 |
+
domain: domainEl.value,
|
| 352 |
+
k: parseInt(kEl.value || "3", 10),
|
| 353 |
+
max_examples: parseInt(maxExamplesEl.value || "5", 10),
|
| 354 |
+
split: splitEl.value || "test"
|
| 355 |
+
};
|
| 356 |
+
|
| 357 |
+
statusEl.textContent = "Running domain evaluation...";
|
| 358 |
+
btn.disabled = true;
|
| 359 |
+
outputEl.textContent = "{}";
|
| 360 |
+
|
| 361 |
+
try {
|
| 362 |
+
const res = await fetch("/run_domain", {
|
| 363 |
+
method: "POST",
|
| 364 |
+
headers: { "Content-Type": "application/json" },
|
| 365 |
+
body: JSON.stringify(payload)
|
| 366 |
+
});
|
| 367 |
+
const data = await res.json();
|
| 368 |
+
if (!res.ok) {
|
| 369 |
+
statusEl.textContent = "Error " + res.status;
|
| 370 |
+
} else {
|
| 371 |
+
statusEl.textContent = "Done.";
|
| 372 |
+
}
|
| 373 |
+
outputEl.textContent = JSON.stringify(data, null, 2);
|
| 374 |
+
} catch (err) {
|
| 375 |
+
statusEl.textContent = "Request failed: " + err;
|
| 376 |
+
outputEl.textContent = "{}";
|
| 377 |
+
} finally {
|
| 378 |
+
btn.disabled = false;
|
| 379 |
+
}
|
| 380 |
+
}
|
| 381 |
+
|
| 382 |
+
async function runExample() {
|
| 383 |
+
const subsetEl = document.getElementById("subset");
|
| 384 |
+
const indexEl = document.getElementById("example_index");
|
| 385 |
+
const kEl = document.getElementById("k_example");
|
| 386 |
+
const splitEl = document.getElementById("split_example");
|
| 387 |
+
const statusEl = document.getElementById("qa_status");
|
| 388 |
+
const outputEl = document.getElementById("qa_output");
|
| 389 |
+
const btn = document.getElementById("qaBtn");
|
| 390 |
+
|
| 391 |
+
const payload = {
|
| 392 |
+
subset: subsetEl.value,
|
| 393 |
+
index: parseInt(indexEl.value || "0", 10),
|
| 394 |
+
k: parseInt(kEl.value || "3", 10),
|
| 395 |
+
split: splitEl.value || "test"
|
| 396 |
+
};
|
| 397 |
+
|
| 398 |
+
statusEl.textContent = "Running single example...";
|
| 399 |
+
btn.disabled = true;
|
| 400 |
+
outputEl.textContent = "{}";
|
| 401 |
+
|
| 402 |
+
try {
|
| 403 |
+
const res = await fetch("/qa_example", {
|
| 404 |
+
method: "POST",
|
| 405 |
+
headers: { "Content-Type": "application/json" },
|
| 406 |
+
body: JSON.stringify(payload)
|
| 407 |
+
});
|
| 408 |
+
const data = await res.json();
|
| 409 |
+
if (!res.ok) {
|
| 410 |
+
statusEl.textContent = "Error " + res.status;
|
| 411 |
+
} else if (data.error) {
|
| 412 |
+
statusEl.textContent = "Backend error: " + data.error;
|
| 413 |
+
} else {
|
| 414 |
+
statusEl.textContent = "Done.";
|
| 415 |
+
}
|
| 416 |
+
outputEl.textContent = JSON.stringify(data, null, 2);
|
| 417 |
+
} catch (err) {
|
| 418 |
+
statusEl.textContent = "Request failed: " + err;
|
| 419 |
+
outputEl.textContent = "{}";
|
| 420 |
+
} finally {
|
| 421 |
+
btn.disabled = false;
|
| 422 |
+
}
|
| 423 |
+
}
|
| 424 |
+
</script>
|
| 425 |
+
</body>
|
| 426 |
+
</html>
|
| 427 |
+
"""
|
| 428 |
+
return HTMLResponse(content=html)
|
docker-compose.yml
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version: "3.9"
|
| 2 |
+
services:
|
| 3 |
+
ragbench-api:
|
| 4 |
+
build: .
|
| 5 |
+
ports:
|
| 6 |
+
- "8000:8000"
|
| 7 |
+
environment:
|
| 8 |
+
HF_TOKEN: "${HF_TOKEN}"
|
| 9 |
+
GROQ_API_KEY: "${GROQ_API_KEY}"
|
| 10 |
+
RAGBENCH_LLM_PROVIDER: "${RAGBENCH_LLM_PROVIDER:-groq}"
|
| 11 |
+
RAGBENCH_GEN_MODEL: "${RAGBENCH_GEN_MODEL:-llama3-8b-8192}"
|
| 12 |
+
RAGBENCH_JUDGE_MODEL: "${RAGBENCH_JUDGE_MODEL:-llama3-70b-8192}"
|
prompts/ragbench_judge_prompt.txt
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
IMPORTANT: Replace this file content with the official JSON-format judge prompt
|
| 2 |
+
from the RAGBench paper (Appendix 9.4). Keep the placeholders:
|
| 3 |
+
{documents}
|
| 4 |
+
{question}
|
| 5 |
+
{answer}
|
| 6 |
+
exactly as they are used in their template.
|
ragbench_eval/__init__.py
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
__all__ = []
|
ragbench_eval/config.py
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
from dotenv import load_dotenv
|
| 3 |
+
|
| 4 |
+
load_dotenv()
|
| 5 |
+
|
| 6 |
+
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 7 |
+
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
|
| 8 |
+
|
| 9 |
+
LLM_PROVIDER = os.getenv("RAGBENCH_LLM_PROVIDER", "groq") # "groq" or "hf"
|
| 10 |
+
GEN_MODEL = os.getenv("RAGBENCH_GEN_MODEL", "llama3-8b-8192")
|
| 11 |
+
JUDGE_MODEL = os.getenv("RAGBENCH_JUDGE_MODEL", "llama3-70b-8192")
|
| 12 |
+
|
| 13 |
+
EMBEDDING_MODEL = os.getenv(
|
| 14 |
+
"RAGBENCH_EMBEDDING_MODEL",
|
| 15 |
+
"sentence-transformers/all-MiniLM-L6-v2",
|
| 16 |
+
)
|
| 17 |
+
|
| 18 |
+
RAGBENCH_DATASET = os.getenv("RAGBENCH_DATASET", "galileo-ai/ragbench")
|
| 19 |
+
|
| 20 |
+
DOMAIN_TO_SUBSETS = {
|
| 21 |
+
"biomedical": ["pubmedqa", "covidqa"],
|
| 22 |
+
"general_knowledge": ["hotpotqa", "msmarco", "hagrid", "expertqa"],
|
| 23 |
+
"legal": ["cuad"],
|
| 24 |
+
"customer_support": ["delucionqa", "emanual", "techqa"],
|
| 25 |
+
"finance": ["finqa", "tatqa"],
|
| 26 |
+
}
|
ragbench_eval/generator.py
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import List, Tuple
|
| 2 |
+
from .llm import LLMClient
|
| 3 |
+
from .config import GEN_MODEL
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
def build_context_from_docs(
|
| 7 |
+
docs_sentences: List[List[Tuple[str, str]]]
|
| 8 |
+
) -> str:
|
| 9 |
+
chunks = []
|
| 10 |
+
for doc in docs_sentences:
|
| 11 |
+
text = " ".join(sent for _, sent in doc)
|
| 12 |
+
chunks.append(text)
|
| 13 |
+
return "\n\n".join(chunks)
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class RAGGenerator:
|
| 17 |
+
def __init__(self):
|
| 18 |
+
self.client = LLMClient(GEN_MODEL)
|
| 19 |
+
|
| 20 |
+
def generate(self, question: str, docs_sentences: List[List[Tuple[str, str]]]) -> str: # noqa: E501
|
| 21 |
+
context = build_context_from_docs(docs_sentences)
|
| 22 |
+
prompt = (
|
| 23 |
+
"Use the following pieces of context to answer the question.\n\n"
|
| 24 |
+
f"{context}\n\n"
|
| 25 |
+
f"Question: {question}\n\n"
|
| 26 |
+
"Answer:"
|
| 27 |
+
)
|
| 28 |
+
messages = [
|
| 29 |
+
{"role": "system", "content": "You are a precise, grounded QA assistant."}, # noqa: E501
|
| 30 |
+
{"role": "user", "content": prompt},
|
| 31 |
+
]
|
| 32 |
+
return self.client.chat(messages)
|
ragbench_eval/judge.py
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import json
|
| 2 |
+
from pathlib import Path
|
| 3 |
+
from typing import Any, Dict, List, Tuple
|
| 4 |
+
|
| 5 |
+
from .llm import LLMClient
|
| 6 |
+
from .config import JUDGE_MODEL
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
def format_docs_with_keys(
|
| 10 |
+
documents_sentences: List[List[Tuple[str, str]]]
|
| 11 |
+
) -> str:
|
| 12 |
+
blocks = []
|
| 13 |
+
for doc in documents_sentences:
|
| 14 |
+
for key, sent in doc:
|
| 15 |
+
blocks.append(f"{key}: {sent}")
|
| 16 |
+
blocks.append("") # blank line
|
| 17 |
+
return "\n".join(blocks).strip()
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
class RAGJudge:
|
| 21 |
+
def __init__(self, prompt_path: str = "prompts/ragbench_judge_prompt.txt"):
|
| 22 |
+
self.client = LLMClient(JUDGE_MODEL)
|
| 23 |
+
self.prompt_template = Path(prompt_path).read_text(encoding="utf-8")
|
| 24 |
+
|
| 25 |
+
def annotate(
|
| 26 |
+
self,
|
| 27 |
+
question: str,
|
| 28 |
+
answer: str,
|
| 29 |
+
docs_sentences: List[List[Tuple[str, str]]],
|
| 30 |
+
) -> Dict[str, Any]:
|
| 31 |
+
docs_block = format_docs_with_keys(docs_sentences)
|
| 32 |
+
prompt = self.prompt_template.format(
|
| 33 |
+
documents=docs_block,
|
| 34 |
+
question=question,
|
| 35 |
+
answer=answer,
|
| 36 |
+
)
|
| 37 |
+
messages = [
|
| 38 |
+
{
|
| 39 |
+
"role": "system",
|
| 40 |
+
"content": "You are an evaluator that outputs STRICT JSON only.",
|
| 41 |
+
},
|
| 42 |
+
{"role": "user", "content": prompt},
|
| 43 |
+
]
|
| 44 |
+
raw = self.client.chat(messages, max_tokens=2048)
|
| 45 |
+
|
| 46 |
+
try:
|
| 47 |
+
data = json.loads(raw)
|
| 48 |
+
except json.JSONDecodeError as e:
|
| 49 |
+
raise ValueError(f"Judge JSON parse error: {e}\nRaw: {raw[:500]}")
|
| 50 |
+
for key in [
|
| 51 |
+
"relevance_explanation",
|
| 52 |
+
"all_relevant_sentence_keys",
|
| 53 |
+
"overall_supported_explanation",
|
| 54 |
+
"overall_supported",
|
| 55 |
+
"sentence_support_information",
|
| 56 |
+
"all_utilized_sentence_keys",
|
| 57 |
+
]:
|
| 58 |
+
if key not in data:
|
| 59 |
+
raise ValueError(f"Missing key in judge output: {key}")
|
| 60 |
+
return data
|
ragbench_eval/llm.py
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import List, Dict
|
| 2 |
+
from .config import LLM_PROVIDER, HF_TOKEN, GROQ_API_KEY
|
| 3 |
+
|
| 4 |
+
from huggingface_hub import InferenceClient
|
| 5 |
+
from groq import Groq
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class LLMClient:
|
| 9 |
+
def __init__(self, model: str, is_chat: bool = True):
|
| 10 |
+
self.provider = LLM_PROVIDER
|
| 11 |
+
self.model = model
|
| 12 |
+
self.is_chat = is_chat
|
| 13 |
+
|
| 14 |
+
if self.provider == "hf":
|
| 15 |
+
if not HF_TOKEN:
|
| 16 |
+
raise RuntimeError("HF_TOKEN is required for HF provider")
|
| 17 |
+
self.client = InferenceClient(token=HF_TOKEN)
|
| 18 |
+
elif self.provider == "groq":
|
| 19 |
+
if not GROQ_API_KEY:
|
| 20 |
+
raise RuntimeError("GROQ_API_KEY is required for Groq provider")
|
| 21 |
+
self.client = Groq(api_key=GROQ_API_KEY)
|
| 22 |
+
else:
|
| 23 |
+
raise ValueError(f"Unsupported provider {self.provider}")
|
| 24 |
+
|
| 25 |
+
def chat(self, messages: List[Dict[str, str]], max_tokens: int = 1024) -> str:
|
| 26 |
+
if self.provider == "hf":
|
| 27 |
+
prompt = ""
|
| 28 |
+
for m in messages:
|
| 29 |
+
role = m.get("role", "user")
|
| 30 |
+
content = m.get("content", "")
|
| 31 |
+
prompt += f"[{role.upper()}]\n{content}\n"
|
| 32 |
+
out = self.client.text_generation(
|
| 33 |
+
prompt,
|
| 34 |
+
model=self.model,
|
| 35 |
+
max_new_tokens=max_tokens,
|
| 36 |
+
temperature=0.2,
|
| 37 |
+
do_sample=False,
|
| 38 |
+
)
|
| 39 |
+
return out
|
| 40 |
+
else:
|
| 41 |
+
resp = self.client.chat.completions.create(
|
| 42 |
+
model=self.model,
|
| 43 |
+
messages=messages,
|
| 44 |
+
max_tokens=max_tokens,
|
| 45 |
+
temperature=0.2,
|
| 46 |
+
)
|
| 47 |
+
return resp.choices[0].message.content
|
ragbench_eval/metrics.py
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any, Dict, List, Tuple
|
| 2 |
+
import numpy as np
|
| 3 |
+
from sklearn.metrics import mean_squared_error, roc_auc_score
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
def _all_sentence_keys(
|
| 7 |
+
docs_sentences: List[List[Tuple[str, str]]]
|
| 8 |
+
) -> List[str]:
|
| 9 |
+
keys: List[str] = []
|
| 10 |
+
for doc in docs_sentences:
|
| 11 |
+
for key, _ in doc:
|
| 12 |
+
keys.append(key)
|
| 13 |
+
return keys
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def trace_from_attributes(
|
| 17 |
+
attrs: Dict[str, Any],
|
| 18 |
+
docs_sentences: List[List[Tuple[str, str]]],
|
| 19 |
+
) -> Dict[str, float]:
|
| 20 |
+
all_keys = _all_sentence_keys(docs_sentences)
|
| 21 |
+
total = len(all_keys)
|
| 22 |
+
if total == 0:
|
| 23 |
+
return {
|
| 24 |
+
"relevance": 0.0,
|
| 25 |
+
"utilization": 0.0,
|
| 26 |
+
"completeness": 0.0,
|
| 27 |
+
"adherence": 0.0,
|
| 28 |
+
}
|
| 29 |
+
|
| 30 |
+
relevant = set(attrs.get("all_relevant_sentence_keys", [])) & set(all_keys)
|
| 31 |
+
utilized = set(attrs.get("all_utilized_sentence_keys", [])) & set(all_keys)
|
| 32 |
+
|
| 33 |
+
relevance = len(relevant) / total if total > 0 else 0.0
|
| 34 |
+
utilization = len(utilized) / total if total > 0 else 0.0
|
| 35 |
+
completeness = (
|
| 36 |
+
len(relevant & utilized) / len(relevant) if relevant else 0.0
|
| 37 |
+
)
|
| 38 |
+
adherence = 1.0 if attrs.get("overall_supported", False) else 0.0
|
| 39 |
+
|
| 40 |
+
return {
|
| 41 |
+
"relevance": float(relevance),
|
| 42 |
+
"utilization": float(utilization),
|
| 43 |
+
"completeness": float(completeness),
|
| 44 |
+
"adherence": float(adherence),
|
| 45 |
+
}
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def compute_rmse_auc(
|
| 49 |
+
y_true_rel: List[float],
|
| 50 |
+
y_pred_rel: List[float],
|
| 51 |
+
y_true_util: List[float],
|
| 52 |
+
y_pred_util: List[float],
|
| 53 |
+
y_true_comp: List[float],
|
| 54 |
+
y_pred_comp: List[float],
|
| 55 |
+
y_true_adh: List[int],
|
| 56 |
+
y_pred_adh: List[float],
|
| 57 |
+
) -> Dict[str, float]:
|
| 58 |
+
metrics = {
|
| 59 |
+
"rmse_relevance": float(
|
| 60 |
+
mean_squared_error(y_true_rel, y_pred_rel, squared=False)
|
| 61 |
+
),
|
| 62 |
+
"rmse_utilization": float(
|
| 63 |
+
mean_squared_error(y_true_util, y_pred_util, squared=False)
|
| 64 |
+
),
|
| 65 |
+
"rmse_completeness": float(
|
| 66 |
+
mean_squared_error(y_true_comp, y_pred_comp, squared=False)
|
| 67 |
+
),
|
| 68 |
+
}
|
| 69 |
+
|
| 70 |
+
if len(set(y_true_adh)) > 1:
|
| 71 |
+
metrics["auroc_adherence"] = float(
|
| 72 |
+
roc_auc_score(y_true_adh, y_pred_adh)
|
| 73 |
+
)
|
| 74 |
+
else:
|
| 75 |
+
metrics["auroc_adherence"] = float("nan")
|
| 76 |
+
|
| 77 |
+
return metrics
|
ragbench_eval/pipeline.py
ADDED
|
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Dict, Any, List, Tuple, Optional
|
| 2 |
+
from datasets import load_dataset
|
| 3 |
+
|
| 4 |
+
from .config import RAGBENCH_DATASET, DOMAIN_TO_SUBSETS
|
| 5 |
+
from .retriever import ExampleRetriever
|
| 6 |
+
from .generator import RAGGenerator
|
| 7 |
+
from .judge import RAGJudge
|
| 8 |
+
from .metrics import trace_from_attributes, compute_rmse_auc
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class RagBenchExperiment:
|
| 12 |
+
def __init__(
|
| 13 |
+
self,
|
| 14 |
+
k: int = 3,
|
| 15 |
+
max_examples: Optional[int] = None,
|
| 16 |
+
split: str = "test",
|
| 17 |
+
):
|
| 18 |
+
self.k = k
|
| 19 |
+
self.max_examples = max_examples
|
| 20 |
+
self.split = split
|
| 21 |
+
|
| 22 |
+
self.retriever = ExampleRetriever()
|
| 23 |
+
self.generator = RAGGenerator()
|
| 24 |
+
self.judge = RAGJudge()
|
| 25 |
+
|
| 26 |
+
def _load_subset(self, subset: str):
|
| 27 |
+
ds = load_dataset(
|
| 28 |
+
RAGBENCH_DATASET, subset, split=self.split
|
| 29 |
+
)
|
| 30 |
+
return ds
|
| 31 |
+
|
| 32 |
+
def _to_docs_sentences(self, row) -> List[List[Tuple[str, str]]]:
|
| 33 |
+
docs: List[List[Tuple[str, str]]] = []
|
| 34 |
+
for doc in row["documents_sentences"]:
|
| 35 |
+
docs.append([(k, s) for k, s in doc])
|
| 36 |
+
return docs
|
| 37 |
+
|
| 38 |
+
def run_subset(self, subset: str) -> Dict[str, Any]:
|
| 39 |
+
ds = self._load_subset(subset)
|
| 40 |
+
|
| 41 |
+
y_true_rel: List[float] = []
|
| 42 |
+
y_pred_rel: List[float] = []
|
| 43 |
+
y_true_util: List[float] = []
|
| 44 |
+
y_pred_util: List[float] = []
|
| 45 |
+
y_true_comp: List[float] = []
|
| 46 |
+
y_pred_comp: List[float] = []
|
| 47 |
+
y_true_adh: List[int] = []
|
| 48 |
+
y_pred_adh: List[float] = []
|
| 49 |
+
|
| 50 |
+
for i, row in enumerate(ds):
|
| 51 |
+
if self.max_examples is not None and i >= self.max_examples:
|
| 52 |
+
break
|
| 53 |
+
|
| 54 |
+
question = row["question"]
|
| 55 |
+
docs_sentences_full = self._to_docs_sentences(row)
|
| 56 |
+
|
| 57 |
+
doc_indices = self.retriever.rank_docs(
|
| 58 |
+
question, docs_sentences_full, k=self.k
|
| 59 |
+
)
|
| 60 |
+
selected_docs = [docs_sentences_full[j] for j in doc_indices]
|
| 61 |
+
|
| 62 |
+
answer = self.generator.generate(question, selected_docs)
|
| 63 |
+
|
| 64 |
+
attrs = self.judge.annotate(question, answer, selected_docs)
|
| 65 |
+
|
| 66 |
+
pred = trace_from_attributes(attrs, selected_docs)
|
| 67 |
+
|
| 68 |
+
y_true_rel.append(float(row["relevance_score"]))
|
| 69 |
+
y_true_util.append(float(row["utilization_score"]))
|
| 70 |
+
y_true_comp.append(float(row["completeness_score"]))
|
| 71 |
+
y_true_adh.append(int(row["adherence_score"]))
|
| 72 |
+
|
| 73 |
+
y_pred_rel.append(pred["relevance"])
|
| 74 |
+
y_pred_util.append(pred["utilization"])
|
| 75 |
+
y_pred_comp.append(pred["completeness"])
|
| 76 |
+
y_pred_adh.append(pred["adherence"])
|
| 77 |
+
|
| 78 |
+
metrics = compute_rmse_auc(
|
| 79 |
+
y_true_rel,
|
| 80 |
+
y_pred_rel,
|
| 81 |
+
y_true_util,
|
| 82 |
+
y_pred_util,
|
| 83 |
+
y_true_comp,
|
| 84 |
+
y_pred_comp,
|
| 85 |
+
y_true_adh,
|
| 86 |
+
y_pred_adh,
|
| 87 |
+
)
|
| 88 |
+
|
| 89 |
+
return {
|
| 90 |
+
"subset": subset,
|
| 91 |
+
"n_examples": len(y_true_rel),
|
| 92 |
+
**metrics,
|
| 93 |
+
}
|
| 94 |
+
|
| 95 |
+
def run_domain(self, domain: str) -> Dict[str, Any]:
|
| 96 |
+
subsets = DOMAIN_TO_SUBSETS[domain]
|
| 97 |
+
results = []
|
| 98 |
+
for subset in subsets:
|
| 99 |
+
res = self.run_subset(subset)
|
| 100 |
+
results.append(res)
|
| 101 |
+
return {
|
| 102 |
+
"domain": domain,
|
| 103 |
+
"subsets": results,
|
| 104 |
+
}
|
ragbench_eval/retriever.py
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import List, Tuple
|
| 2 |
+
import numpy as np
|
| 3 |
+
from sentence_transformers import SentenceTransformer
|
| 4 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
| 5 |
+
|
| 6 |
+
from .config import EMBEDDING_MODEL
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class ExampleRetriever:
|
| 10 |
+
"""Ranks the per-example documents in RAGBench by similarity to the question.""" # noqa: E501
|
| 11 |
+
|
| 12 |
+
def __init__(self):
|
| 13 |
+
self.embedder = SentenceTransformer(EMBEDDING_MODEL)
|
| 14 |
+
|
| 15 |
+
def _encode(self, texts: List[str]) -> np.ndarray:
|
| 16 |
+
return self.embedder.encode(texts, show_progress_bar=False)
|
| 17 |
+
|
| 18 |
+
def rank_docs(
|
| 19 |
+
self,
|
| 20 |
+
question: str,
|
| 21 |
+
documents_sentences: List[List[Tuple[str, str]]],
|
| 22 |
+
k: int = 4,
|
| 23 |
+
) -> List[int]:
|
| 24 |
+
doc_texts = [
|
| 25 |
+
" ".join(sent for _, sent in doc) for doc in documents_sentences
|
| 26 |
+
]
|
| 27 |
+
q_emb = self._encode([question])
|
| 28 |
+
d_emb = self._encode(doc_texts)
|
| 29 |
+
|
| 30 |
+
sims = cosine_similarity(q_emb, d_emb)[0]
|
| 31 |
+
topk_idx = np.argsort(sims)[::-1][:k]
|
| 32 |
+
return topk_idx.tolist()
|
requirements.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
datasets==2.21.0
|
| 2 |
+
sentence-transformers==3.0.1
|
| 3 |
+
scikit-learn==1.5.2
|
| 4 |
+
numpy==1.26.4
|
| 5 |
+
pydantic==2.9.2
|
| 6 |
+
fastapi==0.115.5
|
| 7 |
+
uvicorn[standard]==0.32.0
|
| 8 |
+
python-dotenv==1.0.1
|
| 9 |
+
huggingface_hub[inference]==0.26.2
|
| 10 |
+
groq==0.9.0
|
scripts/run_experiment.py
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import json
|
| 3 |
+
from ragbench_eval.pipeline import RagBenchExperiment
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
def main():
|
| 7 |
+
parser = argparse.ArgumentParser()
|
| 8 |
+
parser.add_argument(
|
| 9 |
+
"--domain",
|
| 10 |
+
type=str,
|
| 11 |
+
required=True,
|
| 12 |
+
choices=[
|
| 13 |
+
"biomedical",
|
| 14 |
+
"general_knowledge",
|
| 15 |
+
"legal",
|
| 16 |
+
"customer_support",
|
| 17 |
+
"finance",
|
| 18 |
+
],
|
| 19 |
+
)
|
| 20 |
+
parser.add_argument("--k", type=int, default=3)
|
| 21 |
+
parser.add_argument("--max_examples", type=int, default=50)
|
| 22 |
+
parser.add_argument("--split", type=str, default="test")
|
| 23 |
+
args = parser.parse_args()
|
| 24 |
+
|
| 25 |
+
exp = RagBenchExperiment(
|
| 26 |
+
k=args.k,
|
| 27 |
+
max_examples=args.max_examples,
|
| 28 |
+
split=args.split,
|
| 29 |
+
)
|
| 30 |
+
results = exp.run_domain(args.domain)
|
| 31 |
+
print(json.dumps(results, indent=2))
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
if __name__ == "__main__":
|
| 35 |
+
main()
|