Merge pull request #452 from Scale3-Labs/development

Release: support `invoke_model` and `converse_stream` for bedrock

Author: alizenhom

src/examples/awsbedrock_examples/__init__.py

@@ -1,10 +1,19 @@
-from examples.awsbedrock_examples.converse import use_converse
+from examples.awsbedrock_examples.converse import (
+    use_converse_stream,
+    use_converse,
+    use_invoke_model_anthropic,
+    use_invoke_model_cohere,
+    use_invoke_model_amazon,
+)
 from langtrace_python_sdk import langtrace, with_langtrace_root_span
 
-langtrace.init()
-
 
 class AWSBedrockRunner:
     @with_langtrace_root_span("AWS_Bedrock")
     def run(self):
+
+        use_converse_stream()
         use_converse()
+        use_invoke_model_anthropic()
+        use_invoke_model_cohere()
+        use_invoke_model_amazon()

src/examples/awsbedrock_examples/converse.py

@@ -1,34 +1,174 @@
-import os
 import boto3
+import json
 from langtrace_python_sdk import langtrace
+from dotenv import load_dotenv
+import botocore
+
+load_dotenv()
+langtrace.init(write_spans_to_console=False)
+
+brt = boto3.client("bedrock-runtime", region_name="us-east-1")
+brc = boto3.client("bedrock", region_name="us-east-1")
+
+
+def use_converse_stream():
+    model_id = "anthropic.claude-3-haiku-20240307-v1:0"
+    conversation = [
+        {
+            "role": "user",
+            "content": [{"text": "what is the capital of France?"}],
+        }
+    ]
+
+    try:
+        response = brt.converse_stream(
+            modelId=model_id,
+            messages=conversation,
+            inferenceConfig={"maxTokens": 4096, "temperature": 0},
+            additionalModelRequestFields={"top_k": 250},
+        )
+        # response_text = response["output"]["message"]["content"][0]["text"]
+        print(response)
+
+    except Exception as e:
+        print(f"ERROR: Can't invoke '{model_id}'. Reason: {e}")
+        exit(1)
 
-langtrace.init(api_key=os.environ["LANGTRACE_API_KEY"])
 
 def use_converse():
     model_id = "anthropic.claude-3-haiku-20240307-v1:0"
-    client = boto3.client(
-        "bedrock-runtime",
-        region_name="us-east-1",
-        aws_access_key_id=os.environ["AWS_ACCESS_KEY_ID"],
-        aws_secret_access_key=os.environ["AWS_SECRET_ACCESS_KEY"],
-    )
     conversation = [
         {
             "role": "user",
-            "content": [{"text": "Write a story about a magic backpack."}],
+            "content": [{"text": "what is the capital of France?"}],
         }
     ]
 
     try:
-        response = client.converse(
+        response = brt.converse(
             modelId=model_id,
             messages=conversation,
-            inferenceConfig={"maxTokens":4096,"temperature":0},
-            additionalModelRequestFields={"top_k":250}
+            inferenceConfig={"maxTokens": 4096, "temperature": 0},
+            additionalModelRequestFields={"top_k": 250},
         )
         response_text = response["output"]["message"]["content"][0]["text"]
         print(response_text)
 
-    except (Exception) as e:
+    except Exception as e:
         print(f"ERROR: Can't invoke '{model_id}'. Reason: {e}")
-        exit(1)
+        exit(1)
+
+
+def get_foundation_models():
+    for model in brc.list_foundation_models()["modelSummaries"]:
+        print(model["modelId"])
+
+
+# Invoke Model API
+# Amazon Titan Models
+def use_invoke_model_titan(stream=False):
+    try:
+        prompt_data = "what's the capital of France?"
+        body = json.dumps(
+            {
+                "inputText": prompt_data,
+                "textGenerationConfig": {
+                    "maxTokenCount": 1024,
+                    "topP": 0.95,
+                    "temperature": 0.2,
+                },
+            }
+        )
+        modelId = "amazon.titan-text-express-v1"  # "amazon.titan-tg1-large"
+        accept = "application/json"
+        contentType = "application/json"
+
+        if stream:
+
+            response = brt.invoke_model_with_response_stream(
+                body=body, modelId=modelId, accept=accept, contentType=contentType
+            )
+        else:
+            response = brt.invoke_model(
+                body=body, modelId=modelId, accept=accept, contentType=contentType
+            )
+            response_body = json.loads(response.get("body").read())
+
+    except botocore.exceptions.ClientError as error:
+
+        if error.response["Error"]["Code"] == "AccessDeniedException":
+            print(
+                f"\x1b[41m{error.response['Error']['Message']}\
+                    \nTo troubeshoot this issue please refer to the following resources.\
+                    \nhttps://docs.aws.amazon.com/IAM/latest/UserGuide/troubleshoot_access-denied.html\
+                    \nhttps://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html\x1b[0m\n"
+            )
+
+        else:
+            raise error
+
+
+# Anthropic Models
+def use_invoke_model_anthropic(stream=False):
+    body = json.dumps(
+        {
+            "anthropic_version": "bedrock-2023-05-31",
+            "max_tokens": 1024,
+            "temperature": 0.1,
+            "top_p": 0.9,
+            "messages": [{"role": "user", "content": "Hello, Claude"}],
+        }
+    )
+    modelId = "anthropic.claude-v2"
+    accept = "application/json"
+    contentType = "application/json"
+
+    if stream:
+        response = brt.invoke_model_with_response_stream(body=body, modelId=modelId)
+        stream_response = response.get("body")
+        if stream_response:
+            for event in stream_response:
+                chunk = event.get("chunk")
+                if chunk:
+                    print(json.loads(chunk.get("bytes").decode()))
+
+    else:
+        response = brt.invoke_model(
+            body=body, modelId=modelId, accept=accept, contentType=contentType
+        )
+        response_body = json.loads(response.get("body").read())
+        # text
+        print(response_body.get("completion"))
+
+
+def use_invoke_model_llama():
+    model_id = "meta.llama3-8b-instruct-v1:0"
+    prompt = "What is the capital of France?"
+    max_gen_len = 128
+    temperature = 0.1
+    top_p = 0.9
+
+    # Create request body.
+    body = json.dumps(
+        {
+            "prompt": prompt,
+            "max_gen_len": max_gen_len,
+            "temperature": temperature,
+            "top_p": top_p,
+        }
+    )
+    response = brt.invoke_model(body=body, modelId=model_id)
+
+    response_body = json.loads(response.get("body").read())
+
+    return response_body
+
+
+# print(get_foundation_models())
+def use_invoke_model_cohere():
+    model_id = "cohere.command-r-plus-v1"
+    prompt = "What is the capital of France?"
+    body = json.dumps({"prompt": prompt, "max_tokens": 1024, "temperature": 0.1})
+    response = brt.invoke_model(body=body, modelId=model_id)
+    response_body = json.loads(response.get("body").read())
+    print(response_body)

src/langtrace_python_sdk/constants/instrumentation/aws_bedrock.py

@@ -1,6 +1,10 @@
 from langtrace.trace_attributes import AWSBedrockMethods
 
 APIS = {
+    "INVOKE_MODEL": {
+        "METHOD": "aws_bedrock.invoke_model",
+        "ENDPOINT": "/invoke-model",
+    },
     "CONVERSE": {
         "METHOD": AWSBedrockMethods.CONVERSE.value,
         "ENDPOINT": "/converse",

src/langtrace_python_sdk/instrumentation/aws_bedrock/bedrock_streaming_wrapper.py

@@ -0,0 +1,43 @@
+import json
+from wrapt import ObjectProxy
+
+
+class StreamingWrapper(ObjectProxy):
+    def __init__(
+        self,
+        response,
+        stream_done_callback=None,
+    ):
+        super().__init__(response)
+
+        self._stream_done_callback = stream_done_callback
+        self._accumulating_body = {}
+
+    def __iter__(self):
+        for event in self.__wrapped__:
+            self._process_event(event)
+            yield event
+
+    def _process_event(self, event):
+        chunk = event.get("chunk")
+        if not chunk:
+            return
+
+        decoded_chunk = json.loads(chunk.get("bytes").decode())
+        type = decoded_chunk.get("type")
+
+        if type == "message_start":
+            self._accumulating_body = decoded_chunk.get("message")
+        elif type == "content_block_start":
+            self._accumulating_body["content"].append(
+                decoded_chunk.get("content_block")
+            )
+        elif type == "content_block_delta":
+            self._accumulating_body["content"][-1]["text"] += decoded_chunk.get(
+                "delta"
+            ).get("text")
+        elif type == "message_stop" and self._stream_done_callback:
+            self._accumulating_body["invocation_metrics"] = decoded_chunk.get(
+                "amazon-bedrock-invocationMetrics"
+            )
+            self._stream_done_callback(self._accumulating_body)

src/langtrace_python_sdk/instrumentation/aws_bedrock/instrumentation.py

@@ -22,22 +22,13 @@
 from opentelemetry.trace import get_tracer
 from wrapt import wrap_function_wrapper as _W
 
-from langtrace_python_sdk.instrumentation.aws_bedrock.patch import (
-    converse, converse_stream
-)
+from langtrace_python_sdk.instrumentation.aws_bedrock.patch import patch_aws_bedrock
 
 logging.basicConfig(level=logging.FATAL)
 
-def _patch_client(client, version: str, tracer) -> None:
-
-    # Store original methods
-    original_converse = client.converse
-
-    # Replace with wrapped versions
-    client.converse = converse("aws_bedrock.converse", version, tracer)(original_converse)
 
 class AWSBedrockInstrumentation(BaseInstrumentor):
-    
+
     def instrumentation_dependencies(self) -> Collection[str]:
         return ["boto3 >= 1.35.31"]
 
@@ -46,13 +37,11 @@ def _instrument(self, **kwargs):
         tracer = get_tracer(__name__, "", tracer_provider)
         version = importlib.metadata.version("boto3")
 
-        def wrap_create_client(wrapped, instance, args, kwargs):
-            result = wrapped(*args, **kwargs)
-            if args and args[0] == 'bedrock-runtime':
-                _patch_client(result, version, tracer)
-            return result
-
-        _W("boto3", "client", wrap_create_client)
+        _W(
+            module="boto3",
+            name="client",
+            wrapper=patch_aws_bedrock(tracer, version),
+        )
 
     def _uninstrument(self, **kwargs):
-        pass
+        pass