Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,76 @@
|
|
1 |
-
from fastapi import FastAPI
|
|
|
|
|
2 |
|
3 |
app = FastAPI()
|
4 |
|
5 |
-
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import FastAPI, HTTPException
|
2 |
+
from pydantic import BaseModel
|
3 |
+
import re
|
4 |
|
5 |
app = FastAPI()
|
6 |
|
7 |
+
# Configuration for enabled guardrail checks as part of the request payload.
|
8 |
+
class GuardrailsConfig(BaseModel):
|
9 |
+
factual_consistency: bool = True
|
10 |
+
toxicity: bool = True
|
11 |
+
# Extend with more flags for additional guardrails
|
12 |
+
|
13 |
+
# Request model now includes both the response and the configuration.
|
14 |
+
class CheckRequest(BaseModel):
|
15 |
+
response: str
|
16 |
+
config: GuardrailsConfig = GuardrailsConfig() # Default config if not provided
|
17 |
+
|
18 |
+
class CheckResponse(BaseModel):
|
19 |
+
grounded: bool
|
20 |
+
details: dict
|
21 |
+
|
22 |
+
# A simple result class to hold individual check outcomes.
|
23 |
+
class Result:
|
24 |
+
def __init__(self):
|
25 |
+
self.details = {}
|
26 |
+
|
27 |
+
def add(self, rule_name: str, passed: bool):
|
28 |
+
self.details[rule_name] = passed
|
29 |
+
|
30 |
+
def grounded(self) -> bool:
|
31 |
+
# The response is considered "grounded" if all enabled rules pass.
|
32 |
+
return all(self.details.values())
|
33 |
+
|
34 |
+
# Define guardrail rule classes.
|
35 |
+
class FactualConsistencyRule:
|
36 |
+
name = "FactualConsistency"
|
37 |
+
|
38 |
+
def check(self, response_text: str) -> bool:
|
39 |
+
# For demonstration: pass if the response contains the word "fact".
|
40 |
+
return "fact" in response_text.lower()
|
41 |
+
|
42 |
+
class ToxicityRule:
|
43 |
+
name = "Toxicity"
|
44 |
+
|
45 |
+
def check(self, response_text: str) -> bool:
|
46 |
+
# For demonstration: fail if negative words like "hate" or "kill" are found.
|
47 |
+
return not re.search(r"(hate|kill)", response_text, re.IGNORECASE)
|
48 |
+
|
49 |
+
# Manager class to load and execute the enabled guardrail rules.
|
50 |
+
class GuardrailsManager:
|
51 |
+
def __init__(self, config: GuardrailsConfig):
|
52 |
+
self.config = config
|
53 |
+
self.rules = self.load_rules()
|
54 |
+
|
55 |
+
def load_rules(self):
|
56 |
+
rules = []
|
57 |
+
if self.config.factual_consistency:
|
58 |
+
rules.append(FactualConsistencyRule())
|
59 |
+
if self.config.toxicity:
|
60 |
+
rules.append(ToxicityRule())
|
61 |
+
# Add additional rules based on configuration here.
|
62 |
+
return rules
|
63 |
+
|
64 |
+
def check(self, response_text: str) -> Result:
|
65 |
+
result = Result()
|
66 |
+
for rule in self.rules:
|
67 |
+
rule_result = rule.check(response_text)
|
68 |
+
result.add(rule.name, rule_result)
|
69 |
+
return result
|
70 |
+
|
71 |
+
# Define the POST endpoint for guardrail checking.
|
72 |
+
@app.post("/guardrails/check", response_model=CheckResponse)
|
73 |
+
async def check_guardrails(request: CheckRequest):
|
74 |
+
manager = GuardrailsManager(request.config)
|
75 |
+
result = manager.check(request.response)
|
76 |
+
return CheckResponse(grounded=result.grounded(), details=result.details)
|