Merge pull request #75 from guardrails-ai/dtam/feature/fast_api

migrate from flask to fast api for uvicorn and asgi support

Dockerfile

@@ -45,7 +45,7 @@ COPY . .
 EXPOSE 8000
 
 # This is our start command; yours might be different.
-# The guardrails-api is a standard Flask application.
-# You can use whatever production server you want that support Flask.
+# The guardrails-api is a standard FastAPI application.
+# You can use whatever production server you want that support FastAPI.
 # Here we use gunicorn
 CMD gunicorn --bind 0.0.0.0:8000 --timeout=90 --workers=2 'guardrails_api.app:create_app(".env", "sample-config.py")'

guardrails_api/api/guards.py

@@ -0,0 +1,324 @@
+import json
+import os
+import inspect
+from typing import Any, Dict, Optional
+from fastapi import HTTPException, Request, APIRouter
+from fastapi.responses import JSONResponse, StreamingResponse
+from urllib.parse import unquote_plus
+from guardrails import AsyncGuard, Guard
+from guardrails.classes import ValidationOutcome
+from opentelemetry.trace import Span
+from guardrails_api_client import Guard as GuardStruct
+from guardrails_api.clients.cache_client import CacheClient
+from guardrails_api.clients.memory_guard_client import MemoryGuardClient
+from guardrails_api.clients.pg_guard_client import PGGuardClient
+from guardrails_api.clients.postgres_client import postgres_is_enabled
+from guardrails_api.utils.get_llm_callable import get_llm_callable
+from guardrails_api.utils.openai import (
+    outcome_to_chat_completion,
+    outcome_to_stream_response,
+)
+from guardrails_api.utils.handle_error import handle_error
+from string import Template
+
+# if no pg_host is set, use in memory guards
+if postgres_is_enabled():
+    guard_client = PGGuardClient()
+else:
+    guard_client = MemoryGuardClient()
+    # Will be defined at runtime
+    import config  # noqa
+
+    exports = config.__dir__()
+    for export_name in exports:
+        export = getattr(config, export_name)
+        is_guard = isinstance(export, Guard)
+        if is_guard:
+            guard_client.create_guard(export)
+
+cache_client = CacheClient()
+
+cache_client.initialize()
+
+router = APIRouter()
+
+
+@router.get("/guards")
+@handle_error
+async def get_guards():
+    guards = guard_client.get_guards()
+    return [g.to_dict() for g in guards]
+
+
+@router.post("/guards")
+@handle_error
+async def create_guard(guard: GuardStruct):
+    if not postgres_is_enabled():
+        raise HTTPException(
+            status_code=501,
+            detail="Not Implemented POST /guards is not implemented for in-memory guards.",
+        )
+    new_guard = guard_client.create_guard(guard)
+    return new_guard.to_dict()
+
+
+@router.get("/guards/{guard_name}")
+@handle_error
+async def get_guard(guard_name: str, asOf: Optional[str] = None):
+    decoded_guard_name = unquote_plus(guard_name)
+    guard = guard_client.get_guard(decoded_guard_name, asOf)
+    if guard is None:
+        raise HTTPException(
+            status_code=404,
+            detail=f"A Guard with the name {decoded_guard_name} does not exist!",
+        )
+    return guard.to_dict()
+
+
+@router.put("/guards/{guard_name}")
+@handle_error
+async def update_guard(guard_name: str, guard: GuardStruct):
+    if not postgres_is_enabled():
+        raise HTTPException(
+            status_code=501,
+            detail="PUT /<guard_name> is not implemented for in-memory guards.",
+        )
+    decoded_guard_name = unquote_plus(guard_name)
+    updated_guard = guard_client.upsert_guard(decoded_guard_name, guard)
+    return updated_guard.to_dict()
+
+
+@router.delete("/guards/{guard_name}")
+@handle_error
+async def delete_guard(guard_name: str):
+    if not postgres_is_enabled():
+        raise HTTPException(
+            status_code=501,
+            detail="DELETE /<guard_name> is not implemented for in-memory guards.",
+        )
+    decoded_guard_name = unquote_plus(guard_name)
+    guard = guard_client.delete_guard(decoded_guard_name)
+    return guard.to_dict()
+
+
+@router.post("/guards/{guard_name}/openai/v1/chat/completions")
+@handle_error
+async def openai_v1_chat_completions(guard_name: str, request: Request):
+    payload = await request.json()
+    decoded_guard_name = unquote_plus(guard_name)
+    guard_struct = guard_client.get_guard(decoded_guard_name)
+    if guard_struct is None:
+        raise HTTPException(
+            status_code=404,
+            detail=f"A Guard with the name {decoded_guard_name} does not exist!",
+        )
+
+    guard = (
+        Guard.from_dict(guard_struct.to_dict())
+        if not isinstance(guard_struct, Guard)
+        else guard_struct
+    )
+    stream = payload.get("stream", False)
+    has_tool_gd_tool_call = any(
+        tool.get("function", {}).get("name") == "gd_response_tool"
+        for tool in payload.get("tools", [])
+    )
+
+    if not stream:
+        validation_outcome: ValidationOutcome = guard(num_reasks=0, **payload)
+        llm_response = guard.history.last.iterations.last.outputs.llm_response_info
+        result = outcome_to_chat_completion(
+            validation_outcome=validation_outcome,
+            llm_response=llm_response,
+            has_tool_gd_tool_call=has_tool_gd_tool_call,
+        )
+        return JSONResponse(content=result)
+    else:
+
+        async def openai_streamer():
+            guard_stream = guard(num_reasks=0, **payload)
+            for result in guard_stream:
+                chunk = json.dumps(
+                    outcome_to_stream_response(validation_outcome=result)
+                )
+                yield f"data: {chunk}\n\n"
+            yield "\n"
+
+        return StreamingResponse(openai_streamer(), media_type="text/event-stream")
+
+
+@router.post("/guards/{guard_name}/validate")
+@handle_error
+async def validate(guard_name: str, request: Request):
+    payload = await request.json()
+    openai_api_key = request.headers.get(
+        "x-openai-api-key", os.environ.get("OPENAI_API_KEY")
+    )
+    decoded_guard_name = unquote_plus(guard_name)
+    guard_struct = guard_client.get_guard(decoded_guard_name)
+
+    llm_output = payload.pop("llmOutput", None)
+    num_reasks = payload.pop("numReasks", None)
+    prompt_params = payload.pop("promptParams", {})
+    llm_api = payload.pop("llmApi", None)
+    args = payload.pop("args", [])
+    stream = payload.pop("stream", False)
+
+    payload["api_key"] = payload.get("api_key", openai_api_key)
+
+    if llm_api is not None:
+        llm_api = get_llm_callable(llm_api)
+        if openai_api_key is None:
+            raise HTTPException(
+                status_code=400,
+                detail="Cannot perform calls to OpenAI without an api key.",
+            )
+
+    guard = guard_struct
+    is_async = inspect.iscoroutinefunction(llm_api)
+
+    if not isinstance(guard_struct, Guard):
+        if is_async:
+            guard = AsyncGuard.from_dict(guard_struct.to_dict())
+        else:
+            guard: Guard = Guard.from_dict(guard_struct.to_dict())
+    elif is_async:
+        guard: Guard = AsyncGuard.from_dict(guard_struct.to_dict())
+
+    if llm_api is None and num_reasks and num_reasks > 1:
+        raise HTTPException(
+            status_code=400,
+            detail="Cannot perform re-asks without an LLM API. Specify llm_api when calling guard(...).",
+        )
+
+    if llm_output is not None:
+        if stream:
+            raise HTTPException(
+                status_code=400, detail="Streaming is not supported for parse calls!"
+            )
+        result: ValidationOutcome = guard.parse(
+            llm_output=llm_output,
+            num_reasks=num_reasks,
+            prompt_params=prompt_params,
+            llm_api=llm_api,
+            **payload,
+        )
+    else:
+        if stream:
+
+            async def guard_streamer():
+                guard_stream = guard(
+                    llm_api=llm_api,
+                    prompt_params=prompt_params,
+                    num_reasks=num_reasks,
+                    stream=stream,
+                    *args,
+                    **payload,
+                )
+                for result in guard_stream:
+                    validation_output = ValidationOutcome.from_guard_history(
+                        guard.history.last
+                    )
+                    yield validation_output, result
+
+            async def validate_streamer(guard_iter):
+                async for validation_output, result in guard_iter:
+                    fragment_dict = result.to_dict()
+                    fragment_dict["error_spans"] = [
+                        json.dumps({"start": x.start, "end": x.end, "reason": x.reason})
+                        for x in guard.error_spans_in_output()
+                    ]
+                    yield json.dumps(fragment_dict) + "\n"
+
+                call = guard.history.last
+                final_validation_output = ValidationOutcome(
+                    callId=call.id,
+                    validation_passed=result.validation_passed,
+                    validated_output=result.validated_output,
+                    history=guard.history,
+                    raw_llm_output=result.raw_llm_output,
+                )
+                final_output_dict = final_validation_output.to_dict()
+                final_output_dict["error_spans"] = [
+                    json.dumps({"start": x.start, "end": x.end, "reason": x.reason})
+                    for x in guard.error_spans_in_output()
+                ]
+                yield json.dumps(final_output_dict) + "\n"
+
+                serialized_history = [call.to_dict() for call in guard.history]
+                cache_key = f"{guard.name}-{final_validation_output.call_id}"
+                await cache_client.set(cache_key, serialized_history, 300)
+
+            return StreamingResponse(
+                validate_streamer(guard_streamer()), media_type="application/json"
+            )
+        else:
+            if inspect.iscoroutinefunction(guard):
+                result: ValidationOutcome = await guard(
+                    llm_api=llm_api,
+                    prompt_params=prompt_params,
+                    num_reasks=num_reasks,
+                    *args,
+                    **payload,
+                )
+            else:
+                result: ValidationOutcome = guard(
+                    llm_api=llm_api,
+                    prompt_params=prompt_params,
+                    num_reasks=num_reasks,
+                    *args,
+                    **payload,
+                )
+
+    serialized_history = [call.to_dict() for call in guard.history]
+    cache_key = f"{guard.name}-{result.call_id}"
+    await cache_client.set(cache_key, serialized_history, 300)
+    return result.to_dict()
+
+
+@router.get("/guards/{guard_name}/history/{call_id}")
+@handle_error
+async def guard_history(guard_name: str, call_id: str):
+    cache_key = f"{guard_name}-{call_id}"
+    return await cache_client.get(cache_key)
+
+
+def collect_telemetry(
+    *,
+    guard: Guard,
+    validate_span: Span,
+    validation_output: ValidationOutcome,
+    prompt_params: Dict[str, Any],
+    result: ValidationOutcome,
+):
+    # Below is all telemetry collection and
+    # should have no impact on what is returned to the user
+    prompt = guard.history.last.inputs.prompt
+    if prompt:
+        prompt = Template(prompt).safe_substitute(**prompt_params)
+        validate_span.set_attribute("prompt", prompt)
+
+    instructions = guard.history.last.inputs.instructions
+    if instructions:
+        instructions = Template(instructions).safe_substitute(**prompt_params)
+        validate_span.set_attribute("instructions", instructions)
+
+    validate_span.set_attribute("validation_status", guard.history.last.status)
+    validate_span.set_attribute("raw_llm_ouput", result.raw_llm_output)
+
+    # Use the serialization from the class instead of re-writing it
+    valid_output: str = (
+        json.dumps(validation_output.validated_output)
+        if isinstance(validation_output.validated_output, dict)
+        else str(validation_output.validated_output)
+    )
+    validate_span.set_attribute("validated_output", valid_output)
+
+    validate_span.set_attribute("tokens_consumed", guard.history.last.tokens_consumed)
+
+    num_of_reasks = (
+        guard.history.last.iterations.length - 1
+        if guard.history.last.iterations.length > 0
+        else 0
+    )
+    validate_span.set_attribute("num_of_reasks", num_of_reasks)