Add Bias Detection Microservice (#659)

* Add Bias Detection Microservice

Signed-off-by: Qun Gao <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Adding Bias Detection Container to CI (#695)

Signed-off-by: Abolfazl Shahbazi <[email protected]>

---------

Signed-off-by: Qun Gao <[email protected]>
Signed-off-by: Abolfazl Shahbazi <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Abolfazl Shahbazi <[email protected]>

.github/workflows/docker/compose/guardrails-compose-cd.yaml

@@ -6,6 +6,10 @@ services:
     build:
       dockerfile: comps/guardrails/pii_detection/Dockerfile
     image: ${REGISTRY:-opea}/guardrails-pii-detection:${TAG:-latest}
+  guardrails-bias-detection:
+    build:
+      dockerfile: comps/guardrails/bias_detection/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-bias-detection:${TAG:-latest}
   guardrails-toxicity-detection:
     build:
       dockerfile: comps/guardrails/toxicity_detection/Dockerfile

.github/workflows/docker/compose/guardrails-compose.yaml

@@ -9,3 +9,13 @@ services:
     build:
       dockerfile: comps/guardrails/llama_guard/langchain/Dockerfile
     image: ${REGISTRY:-opea}/guardrails-tgi:${TAG:-latest}
+
+  guardrails-bias-detection:
+    build:
+      dockerfile: comps/guardrails/bias_detection/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-bias-detection:${TAG:-latest}
+
+  guardrails-toxicity-detection:
+    build:
+      dockerfile: comps/guardrails/toxicity_detection/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-toxicity-detection:${TAG:-latest}

comps/guardrails/README.md

@@ -7,5 +7,6 @@ The Guardrails service enhances the security of LLM-based applications by offeri
 | [Llama Guard](./llama_guard/langchain/README.md)     | Provides guardrails for inputs and outputs to ensure safe interactions                                                   |
 | [PII Detection](./pii_detection/README.md)           | Detects Personally Identifiable Information (PII) and Business Sensitive Information (BSI)                               |
 | [Toxicity Detection](./toxicity_detection/README.md) | Detects Toxic language (rude, disrespectful, or unreasonable language that is likely to make someone leave a discussion) |
+| [Bias Detection](./bias_detection/README.md)         | Detects Biased language (framing bias, epistemological bias, and demographic bias)                                       |
 
 Additional safety-related microservices will be available soon.

comps/guardrails/bias_detection/Dockerfile

@@ -0,0 +1,31 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+FROM langchain/langchain:latest
+
+ENV LANG=C.UTF-8
+
+ARG ARCH="cpu"
+
+RUN apt-get update -y && apt-get install -y --no-install-recommends --fix-missing \
+    libgl1-mesa-glx \
+    libjemalloc-dev
+
+
+RUN useradd -m -s /bin/bash user && \
+    mkdir -p /home/user && \
+    chown -R user /home/user/
+
+USER user
+
+COPY comps /home/user/comps
+
+RUN pip install --no-cache-dir --upgrade pip && \
+    if [ ${ARCH} = "cpu" ]; then pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu; fi && \
+    pip install --no-cache-dir -r /home/user/comps/guardrails/bias_detection/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home/user
+
+WORKDIR /home/user/comps/guardrails/bias_detection/
+
+ENTRYPOINT ["python", "bias_detection.py"]

comps/guardrails/bias_detection/README.md

@@ -0,0 +1,93 @@
+# Bias Detection Microservice
+
+## Introduction
+
+Bias Detection Microservice allows AI Application developers to safeguard user input and LLM output from biased language in a RAG environment. By leveraging a smaller fine-tuned Transformer model for bias classification (e.g. DistilledBERT, RoBERTa, etc.), we maintain a lightweight guardrails microservice without significantly sacrificing performance making it readily deployable on both Intel Gaudi and Xeon.
+
+Bias erodes our collective trust and fuels social conflict. Bias can be defined as inappropriate subjectivity in the form of one of the following:
+
+- Framing bias -- using subjective words or phrases linked with a particular point of view
+- Epistemological bias -- linguistic features that subtly modify the believability of a proposition
+- Demographic bias -- text with presuppositions about particular genders, races, or other demographic categories
+
+## Future Development
+
+- Add a "neutralizing bias" microservice to neutralizing any detected bias in the RAG serving, guarding the RAG usage.
+
+## 🚀1. Start Microservice with Python（Option 1）
+
+### 1.1 Install Requirements
+
+```bash
+pip install -r requirements.txt
+```
+
+### 1.2 Start Bias Detection Microservice with Python Script
+
+```bash
+python bias_detection.py
+```
+
+## 🚀2. Start Microservice with Docker (Option 2)
+
+### 2.1 Prepare bias detection model
+
+export HUGGINGFACEHUB_API_TOKEN=${HP_TOKEN}
+
+### 2.2 Build Docker Image
+
+```bash
+cd ../../../ # back to GenAIComps/ folder
+docker build -t opea/guardrails-bias-detection:latest --build-arg https_proxy=$https_proxy --build-arg http_proxy=$http_proxy -f comps/guardrails/bias_detection/Dockerfile .
+```
+
+### 2.3 Run Docker Container with Microservice
+
+```bash
+docker run -d --rm --runtime=runc --name="guardrails-bias-detection" -p 9092:9092 --ipc=host -e http_proxy=$http_proxy -e https_proxy=$https_proxy -e HUGGINGFACEHUB_API_TOKEN=${HUGGINGFACEHUB_API_TOKEN} -e HF_TOKEN=${HUGGINGFACEHUB_API_TOKEN} opea/guardrails-bias-detection:latest
+```
+
+## 🚀3. Get Status of Microservice
+
+```bash
+docker container logs -f guardrails-bias-detection
+```
+
+## 🚀4. Consume Microservice Pre-LLM/Post-LLM
+
+Once microservice starts, users can use examples (bash or python) below to apply bias detection for both user's query (Pre-LLM) or LLM's response (Post-LLM)
+
+**Bash:**
+
+```bash
+curl localhost:9092/v1/bias
+    -X POST
+    -d '{"text":"John McCain exposed as an unprincipled politician"}'
+    -H 'Content-Type: application/json'
+```
+
+Example Output:
+
+```bash
+"\nI'm sorry, but your query or LLM's response is BIASED with an score of 0.74 (0-1)!!!\n"
+```
+
+**Python Script:**
+
+```python
+import requests
+import json
+
+proxies = {"http": ""}
+url = "http://localhost:9092/v1/bias"
+data = {"text": "John McCain exposed as an unprincipled politician"}
+
+
+try:
+    resp = requests.post(url=url, data=data, proxies=proxies)
+    print(resp.text)
+    resp.raise_for_status()  # Raise an exception for unsuccessful HTTP status codes
+    print("Request successful!")
+except requests.exceptions.RequestException as e:
+    print("An error occurred:", e)
+```

comps/guardrails/bias_detection/bias_detection.py

@@ -0,0 +1,31 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+from transformers import pipeline
+
+from comps import ServiceType, TextDoc, opea_microservices, register_microservice
+
+
+@register_microservice(
+    name="opea_service@bias_detection",
+    service_type=ServiceType.GUARDRAIL,
+    endpoint="/v1/bias",
+    host="0.0.0.0",
+    port=9092,
+    input_datatype=TextDoc,
+    output_datatype=TextDoc,
+)
+def llm_generate(input: TextDoc):
+    input_text = input.text
+    toxic = bias_pipeline(input_text)
+    print("done")
+    if toxic[0]["label"] == "BIASED":
+        return TextDoc(text="Violated policies: bias, please check your input.", downstream_black_list=[".*"])
+    else:
+        return TextDoc(text=input_text)
+
+
+if __name__ == "__main__":
+    model = "valurank/distilroberta-bias"
+    bias_pipeline = pipeline("text-classification", model=model, tokenizer=model)
+    opea_microservices["opea_service@bias_detection"].start()

comps/guardrails/bias_detection/requirements.txt

@@ -0,0 +1,15 @@
+aiohttp
+docarray[full]
+fastapi
+httpx
+huggingface_hub
+langchain-community
+langchain-huggingface
+opentelemetry-api
+opentelemetry-exporter-otlp
+opentelemetry-sdk
+prometheus-fastapi-instrumentator
+pyyaml
+requests
+shortuuid
+uvicorn