Prediction Guard Guardrails components (opea-project#677)

* added files for PG guardrails components

Signed-off-by: sharanshirodkar7 <[email protected]>

* Fix pre-commit issues: end-of-file, requirements.txt, trailing whitespace, imports, and formatting

Signed-off-by: sharanshirodkar7 <[email protected]>

* added package

Signed-off-by: sharanshirodkar7 <[email protected]>

* added package

Signed-off-by: sharanshirodkar7 <[email protected]>

* added package

Signed-off-by: sharanshirodkar7 <[email protected]>

* file structure updated to latest

Signed-off-by: sharanshirodkar7 <[email protected]>

* Fix pre-commit issues: end-of-file, requirements.txt, trailing whitespace, imports, and formatting

Signed-off-by: sharanshirodkar7 <[email protected]>

* Add .DS_Store to .gitignore

Signed-off-by: sharanshirodkar7 <[email protected]>

* updated readme,requirements+changes based on feedback

Signed-off-by: sharanshirodkar7 <[email protected]>

* references classes in init.py

Signed-off-by: sharanshirodkar7 <[email protected]>

* fix readme link error

Signed-off-by: sharanshirodkar7 <[email protected]>

* fix readme link error

Signed-off-by: sharanshirodkar7 <[email protected]>

* fix readme link error

Signed-off-by: sharanshirodkar7 <[email protected]>

* fix readme link error

Signed-off-by: sharanshirodkar7 <[email protected]>

* fix readme link error

Signed-off-by: sharanshirodkar7 <[email protected]>

* fix readme link error

Signed-off-by: sharanshirodkar7 <[email protected]>

* Fix pre-commit issues: end-of-file, requirements.txt, trailing whitespace, imports, and formatting

Signed-off-by: sharanshirodkar7 <[email protected]>

* removed duplicatesa

Signed-off-by: sharanshirodkar7 <[email protected]>

* removed added readme content

Signed-off-by: sharanshirodkar7 <[email protected]>

---------

Signed-off-by: sharanshirodkar7 <[email protected]>

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

@@ -14,3 +14,19 @@ services:
     build:
       dockerfile: comps/guardrails/toxicity_detection/Dockerfile
     image: ${REGISTRY:-opea}/guardrails-toxicity-detection:${TAG:-latest}
+  guardrails-pii-detection-predictionguard:
+    build:
+      dockerfile: comps/guardrails/pii_detection/predictionguard/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-pii-predictionguard:${TAG:-latest}
+  guardrails-toxicity-detection-predictionguard:
+    build:
+      dockerfile: comps/guardrails/toxicity_detection/predictionguard/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-toxicity-predictionguard:${TAG:-latest}
+  guardrails-factuality-predictionguard:
+    build:
+      dockerfile: comps/guardrails/factuality/predictionguard/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-factuality-predictionguard:${TAG:-latest}
+  guardrails-injection-predictionguard:
+    build:
+      dockerfile: comps/guardrails/prompt_injection/predictionguard/Dockerfile
+    image: ${REGISTRY:-opea}/guardrails-injection-predictionguard:${TAG:-latest}

.gitignore

@@ -1,2 +1,3 @@
 __pycache__
 *.egg-info/
+.DS_Store

comps/__init__.py

@@ -29,6 +29,10 @@
     TextImageDoc,
     MultimodalDoc,
     EmbedMultimodalDoc,
+    FactualityDoc,
+    ScoreDoc,
+    PIIRequestDoc,
+    PIIResponseDoc,
 )
 
 # Constants

comps/cores/proto/docarray.py

@@ -20,6 +20,26 @@ class TextDoc(BaseDoc, TopologyInfo):
     text: str = None
 
 
+class FactualityDoc(BaseDoc):
+    reference: str
+    text: str
+
+
+class ScoreDoc(BaseDoc):
+    score: float
+
+
+class PIIRequestDoc(BaseDoc):
+    prompt: str
+    replace: Optional[bool] = False
+    replace_method: Optional[str] = "random"
+
+
+class PIIResponseDoc(BaseDoc):
+    detected_pii: Optional[List[dict]] = None
+    new_prompt: Optional[str] = None
+
+
 class MetadataTextDoc(TextDoc):
     metadata: Optional[Dict[str, Any]] = Field(
         description="This encloses all metadata associated with the textdoc.",

comps/guardrails/factuality/predictionguard/Dockerfile

@@ -0,0 +1,15 @@
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identitier: Apache-2.0
+
+FROM python:3.11-slim
+
+COPY comps /home/comps
+
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r /home/comps/guardrails/factuality/predictionguard/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home
+
+WORKDIR /home/comps/guardrails/factuality/predictionguard
+
+ENTRYPOINT ["python", "factuality_predictionguard.py" ]

comps/guardrails/factuality/predictionguard/README.md

@@ -0,0 +1,39 @@
+# Factuality Check Prediction Guard Microservice
+
+[Prediction Guard](https://docs.predictionguard.com) allows you to utilize hosted open access LLMs, LVMs, and embedding functionality with seamlessly integrated safeguards. In addition to providing a scalable access to open models, Prediction Guard allows you to configure factual consistency checks, toxicity filters, PII filters, and prompt injection blocking. Join the [Prediction Guard Discord channel](https://discord.gg/TFHgnhAFKd) and request an API key to get started.
+
+Checking for factual consistency can help to ensure that any LLM hallucinations are being found before being returned to a user. This microservice allows the user to compare two text passages (`reference` and `text`). The output will be a float number from 0.0 to 1.0 (with closer to 1.0 indicating more factual consistency between `reference` and `text`).
+
+# 🚀 Start Microservice with Docker
+
+## Setup Environment Variables
+
+Setup the following environment variables first
+
+```bash
+export PREDICTIONGUARD_API_KEY=${your_predictionguard_api_key}
+```
+
+## Build Docker Images
+
+```bash
+cd ../../../../
+docker build -t opea/factuality-predictionguard:latest -f comps/guardrails/factuality/predictionguard/Dockerfile .
+```
+
+## Start Service
+
+```bash
+docker run -d --name="guardrails-factuality-predictionguard" -p 9075:9075 -e PREDICTIONGUARD_API_KEY=$PREDICTIONGUARD_API_KEY opea/guardrails-factuality-predictionguard:latest
+```
+
+# 🚀 Consume Factuality Check Service
+
+```bash
+curl -X POST http://localhost:9075/v1/factuality \
+    -H 'Content-Type: application/json' \
+    -d '{
+      "reference": "The sky is blue.",
+      "text": "The sky is green."
+    }'
+```

comps/guardrails/factuality/predictionguard/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identifier: Apache-2.0

comps/guardrails/factuality/predictionguard/docker_compose_factuality.yaml

@@ -0,0 +1,20 @@
+# Copyright (C) 2024 Prediction Guard, Inc
+# SPDX-License-Identifier: Apache-2.0
+
+services:
+  factuality:
+    image: opea/guardrails-factuality-predictionguard:latest
+    container_name: guardrails-factuality-predictionguard
+    ports:
+      - "9075:9075"
+    ipc: host
+    environment:
+      no_proxy: ${no_proxy}
+      http_proxy: ${http_proxy}
+      https_proxy: ${https_proxy}
+      PREDICTIONGUARD_API_KEY: ${PREDICTIONGUARD_API_KEY}
+    restart: unless-stopped
+
+networks:
+  default:
+    driver: bridge

comps/guardrails/factuality/predictionguard/factuality_predictionguard.py

@@ -0,0 +1,47 @@
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identified: Apache-2.0
+
+
+import time
+
+from docarray import BaseDoc
+from predictionguard import PredictionGuard
+
+from comps import (
+    FactualityDoc,
+    ScoreDoc,
+    ServiceType,
+    opea_microservices,
+    register_microservice,
+    register_statistics,
+    statistics_dict,
+)
+
+
+@register_microservice(
+    name="opea_service@factuality_predictionguard",
+    service_type=ServiceType.GUARDRAIL,
+    endpoint="/v1/factuality",
+    host="0.0.0.0",
+    port=9075,
+    input_datatype=FactualityDoc,
+    output_datatype=ScoreDoc,
+)
+@register_statistics(names=["opea_service@factuality_predictionguard"])
+def factuality_guard(input: FactualityDoc) -> ScoreDoc:
+    start = time.time()
+
+    client = PredictionGuard()
+
+    reference = input.reference
+    text = input.text
+
+    result = client.factuality.check(reference=reference, text=text)
+
+    statistics_dict["opea_service@factuality_predictionguard"].append_latency(time.time() - start, None)
+    return ScoreDoc(score=result["checks"][0]["score"])
+
+
+if __name__ == "__main__":
+    print("Prediction Guard Factuality initialized.")
+    opea_microservices["opea_service@factuality_predictionguard"].start()

comps/guardrails/factuality/predictionguard/requirements.txt

@@ -0,0 +1,13 @@
+aiohttp
+docarray
+fastapi
+huggingface_hub
+opentelemetry-api
+opentelemetry-exporter-otlp
+opentelemetry-sdk
+Pillow
+predictionguard
+prometheus-fastapi-instrumentator
+shortuuid
+transformers
+uvicorn

comps/guardrails/pii_detection/predictionguard/Dockerfile

@@ -0,0 +1,16 @@
+
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identitier: Apache-2.0
+
+FROM python:3.11-slim
+
+COPY comps /home/comps
+
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r /home/comps/guardrails/pii_detection/predictionguard/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home
+
+WORKDIR /home/comps/guardrails/pii_detection/predictionguard
+
+ENTRYPOINT ["python", "pii_predictionguard.py" ]

comps/guardrails/pii_detection/predictionguard/README.md

@@ -0,0 +1,50 @@
+# PII Detection Prediction Guard Microservice
+
+[Prediction Guard](https://docs.predictionguard.com) allows you to utilize hosted open access LLMs, LVMs, and embedding functionality with seamlessly integrated safeguards. In addition to providing a scalable access to open models, Prediction Guard allows you to configure factual consistency checks, toxicity filters, PII filters, and prompt injection blocking. Join the [Prediction Guard Discord channel](https://discord.gg/TFHgnhAFKd) and request an API key to get started.
+
+Detecting Personal Identifiable Information (PII) is important in ensuring that users aren't sending out private data to LLMs. This service allows you to configurably:
+
+1. Detect PII
+2. Replace PII (with "faked" information)
+3. Mask PII (with placeholders)
+
+# 🚀 Start Microservice with Docker
+
+## Setup Environment Variables
+
+Setup the following environment variables first
+
+```bash
+export PREDICTIONGUARD_API_KEY=${your_predictionguard_api_key}
+```
+
+## Build Docker Images
+
+```bash
+cd ../../../../
+docker build -t opea/guardrails-pii-predictionguard:latest -f comps/guardrails/pii_detection/predictionguard/Dockerfile .
+```
+
+## Start Service
+
+```bash
+docker run -d --name="guardrails-pii-predictionguard" -p 9080:9080 -e PREDICTIONGUARD_API_KEY=$PREDICTIONGUARD_API_KEY opea/guardrails-pii-predictionguard:latest
+```
+
+# 🚀 Consume PII Detection Service
+
+```bash
+curl -X POST http://localhost:9080/v1/pii \
+    -H 'Content-Type: application/json' \
+    -d '{
+      "prompt": "My name is John Doe and my phone number is 555-555-5555.",
+      "replace": true,
+      "replace_method": "random"
+    }'
+```
+
+API parameters:
+
+- `prompt` (string, required): The text in which you want to detect PII (typically the prompt that you anticipate sending to an LLM)
+- `replace` (boolean, optional, default is `false`): `true` if you want to replace the detected PII in the `prompt`
+- `replace_method` (string, optional, default is `random`): The method you want to use to replace PII (set to either `random`, `fake`, `category`, `mask`)

comps/guardrails/pii_detection/predictionguard/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identifier: Apache-2.0

comps/guardrails/pii_detection/predictionguard/docker_compose_pii.yaml

@@ -0,0 +1,20 @@
+# Copyright (C) 2024 Prediction Guard, Inc
+# SPDX-License-Identifier: Apache-2.0
+
+services:
+  pii:
+    image: opea/guardrails-pii-predictionguard:latest
+    container_name: pii-predictionguard
+    ports:
+      - "9080:9080"
+    ipc: host
+    environment:
+      no_proxy: ${no_proxy}
+      http_proxy: ${http_proxy}
+      https_proxy: ${https_proxy}
+      PREDICTIONGUARD_API_KEY: ${PREDICTIONGUARD_API_KEY}
+    restart: unless-stopped
+
+networks:
+  default:
+    driver: bridge

comps/guardrails/pii_detection/predictionguard/pii_predictionguard.py

@@ -0,0 +1,54 @@
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identified: Apache-2.0
+
+
+import json
+import time
+from typing import List, Optional
+
+from docarray import BaseDoc
+from predictionguard import PredictionGuard
+
+from comps import (
+    PIIRequestDoc,
+    PIIResponseDoc,
+    ServiceType,
+    opea_microservices,
+    register_microservice,
+    register_statistics,
+    statistics_dict,
+)
+
+
+@register_microservice(
+    name="opea_service@pii_predictionguard",
+    service_type=ServiceType.GUARDRAIL,
+    endpoint="/v1/pii",
+    host="0.0.0.0",
+    port=9080,
+    input_datatype=PIIRequestDoc,
+    output_datatype=PIIResponseDoc,
+)
+@register_statistics(names=["opea_service@pii_predictionguard"])
+def pii_guard(input: PIIRequestDoc) -> PIIResponseDoc:
+    start = time.time()
+
+    client = PredictionGuard()
+
+    prompt = input.prompt
+    replace = input.replace
+    replace_method = input.replace_method
+
+    result = client.pii.check(prompt=prompt, replace=replace, replace_method=replace_method)
+
+    statistics_dict["opea_service@pii_predictionguard"].append_latency(time.time() - start, None)
+    if "new_prompt" in result["checks"][0].keys():
+        return PIIResponseDoc(new_prompt=result["checks"][0]["new_prompt"])
+    elif "pii_types_and_positions" in result["checks"][0].keys():
+        detected_pii = json.loads(result["checks"][0]["pii_types_and_positions"])
+        return PIIResponseDoc(detected_pii=detected_pii)
+
+
+if __name__ == "__main__":
+    print("Prediction Guard PII Detection initialized.")
+    opea_microservices["opea_service@pii_predictionguard"].start()

comps/guardrails/pii_detection/predictionguard/requirements.txt

@@ -0,0 +1,13 @@
+aiohttp
+docarray
+fastapi
+huggingface_hub
+opentelemetry-api
+opentelemetry-exporter-otlp
+opentelemetry-sdk
+Pillow
+predictionguard
+prometheus-fastapi-instrumentator
+shortuuid
+transformers
+uvicorn

comps/guardrails/prompt_injection/predictionguard/Dockerfile

@@ -0,0 +1,16 @@
+
+# Copyright (C) 2024 Prediction Guard, Inc.
+# SPDX-License-Identitier: Apache-2.0
+
+FROM python:3.11-slim
+
+COPY comps /home/comps
+
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r /home/comps/guardrails/prompt_injection/predictionguard/requirements.txt
+
+ENV PYTHONPATH=$PYTHONPATH:/home
+
+WORKDIR /home/comps/guardrails/prompt_injection/predictionguard
+
+ENTRYPOINT ["python", "injection_predictionguard.py" ]