changed to explicit map reduce function

.gitignore

@@ -75,3 +75,5 @@ __pycache__/
 
 # Customized ENV files
 .env*
+.venv/
+output.json

cdk/examples/generative_ai_service/lib/gen_ai_service_stack.py

@@ -56,7 +56,11 @@ def __init__(
         fargate_task_image = ApplicationLoadBalancedTaskImageOptions(
             container_name=self.stack_props.container_name,
             # build container image from local folder
-            image=ContainerImage.from_asset("web-app", platform=Platform.LINUX_AMD64),
+            # image=ContainerImage.from_asset("web-app", platform=Platform.LINUX_AMD64),
+            # load pre-built image from public repository
+            image=ContainerImage.from_registry(
+                self.stack_props.container_image
+            ),
             environment={'region': self.stack_props.aws_region},
             container_port=self.stack_props.container_port,
             execution_role=self.ecs_task_execution_role,

cdk/examples/generative_ai_service/sample.env

@@ -19,6 +19,7 @@ az_count="3"
 
 # Gen AI Service Props
 container_name="web-container"
+container_image="public.ecr.aws/g4t0l7p7/sc-genai:latest"
 container_port="80"
 task_cpu="2048"
 task_memory="8192"

cdk/examples/generative_ai_service/web-app/pages/1_text_generation.py

@@ -0,0 +1,71 @@
+import streamlit as st
+import requests
+import json
+import time
+import boto3
+
+st.set_page_config(
+    page_title="text generation",
+    layout="wide",
+    page_icon=":technologist:"
+)
+
+st.header("Generative AI Demo - Text Generation :books:")
+st.caption("Using FLAN-T5-XL model from Hugging Face")
+
+def get_parameter(name):
+    ssm = boto3.client('ssm')
+    param = ssm.get_parameter(Name=name,WithDecryption=True)
+    return param['Parameter']['Value']
+
+runtime = boto3.client("runtime.sagemaker")
+
+conversation = """Customers were very excited about the wireless charging feature, but the launch has not lived up to their expectations. The phones are not reliably charging and that is frustrating since it is such a fundamental aspect of any electronic device."""
+
+with st.spinner("Retrieving configurations..."):
+    all_configs_loaded = False
+
+    while not all_configs_loaded:
+        try:
+            # Retrive SageMaker Endpoint name from Parameter Store
+            sm_endpoint = get_parameter("txt2txt_sm_endpoint")
+            all_configs_loaded = True
+        except:
+            time.sleep(5)
+
+    endpoint_name = st.sidebar.text_input("SageMaker Endpoint Name:",sm_endpoint)
+
+    context = st.text_area("Input Context:", conversation, height=300)
+
+    query = st.text_area("Input Query:", "Are customers happy?")
+    st.caption("e.g., write a summary")
+
+    if st.button("Generate Response", key=query):
+        if endpoint_name == "" or query == "":
+            st.error("Please enter a valid endpoint name and prompt!")
+        else:
+            with st.spinner("Wait for it..."):
+                try:
+                    prompt = f"{context}\n{query}"
+                    response = runtime.invoke_endpoint(
+                        EndpointName=endpoint_name,
+                        Body=prompt,
+                        ContentType="application/x-text",
+                    )
+                    response_body = json.loads(response["Body"].read().decode())
+                    generated_text = response_body["generated_text"]
+                    st.write(generated_text)
+
+                except requests.exceptions.ConnectionError as errc:
+                    st.error("Error Connecting:",errc)
+
+                except requests.exceptions.HTTPError as errh:
+                    st.error("Http Error:",errh)
+
+                except requests.exceptions.Timeout as errt:
+                    st.error("Timeout Error:",errt)
+
+                except requests.exceptions.RequestException as err:
+                    st.error("OOps: Something Else",err)
+
+            st.success("Done!")

cdk/examples/generative_ai_service/web-app/pages/llm_langchain.py → cdk/examples/generative_ai_service/web-app/pages/2_llm_langchain.py

@@ -1,6 +1,7 @@
 import json, boto3
 from langchain.prompts import PromptTemplate
 from langchain.llms.sagemaker_endpoint import LLMContentHandler, SagemakerEndpoint
+from langchain import LLMChain
 from langchain.chains.summarize import load_summarize_chain
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 import streamlit as st
@@ -46,17 +47,41 @@ def create_chunks(context):
 
     return docs
 
-def invoke_map_reduce(context, query, endpoint_name, region_name):
+def simple_map_reduce(context, map_query, reduce_query, endpoint_name, region_name):
+    docs = create_chunks(context)
+    content_handler = ContentHandler()
+    llm = SagemakerEndpoint(
+                    endpoint_name = endpoint_name,
+                    region_name= region_name,
+                    content_handler=content_handler
+                )
+
+    map_template = """In the following text find all answers to the question which is delimited by ```. {text}. Provide answer as complete sentence.""" + """```""" + map_query + """```\n"""
+    map_prompt = PromptTemplate(template=map_template,
+                                input_variables=["text"])
+    map_chain = LLMChain(llm=llm,prompt=map_prompt)
+
+    output_list =[]
+    for doc in docs:
+        output = map_chain.run(doc)
+        output_list.append(output)
+
+    reduce_template = """In the following comma separated text list, find all answers to the question which is delimited by ```. {text}. Provide answer as complete sentence.""" + """```""" + reduce_query + """```\n"""
+    reduce_prompt = PromptTemplate(template=reduce_template,
+                                input_variables=["text"])
+    reduce_chain = LLMChain(llm=llm,prompt=reduce_prompt)
+    reduce_output = reduce_chain.run(','.join(output_list))
+    return {"map_answers":output_list, "reduce_answer":reduce_output}
+
+def langchain_map_reduce(context, map_query, reduce_query, endpoint_name, region_name):
     content_handler = ContentHandler()
 
     map_template = """{text}
     Question: {question}
-    Question's answer:"""
+    Answer:"""
     map_prompt_template = PromptTemplate(template=map_template, input_variables=["text","question"])
 
-    combine_template = """{text}
-    Combine above texts and rewrite into summary.
-    """
+    combine_template = """{text} """ + """ Question: """ + reduce_query + """ Answer:"""
 
     combine_prompt_template = PromptTemplate(template=combine_template, input_variables=["text"])
 
@@ -77,7 +102,7 @@ def invoke_map_reduce(context, query, endpoint_name, region_name):
 
     # split pharagraph
     docs = create_chunks(context)
-    summary = summary_chain({"input_documents": docs, "question": query, 'token_max': 10000}, return_only_outputs=True)
+    summary = summary_chain({"input_documents": docs, "question": map_query, 'token_max': 10000}, return_only_outputs=True)
     return summary.get("output_text","**no response found**")
 
 conversation = """
@@ -103,14 +128,16 @@ def invoke_map_reduce(context, query, endpoint_name, region_name):
 
     context = st.text_area("Input Context:", conversation, height=700)
 
-    query = st.text_area("Input Query:", "What are being introduced in detail?")
+    map_query = st.text_area("Map Query:", "What is being introduced in this blog text?")
+
+    reduce_query = st.text_area("Reduce Query:", "What is the common topic?")
 
-    if st.button("Generate Response", key=query):
-        if endpoint_name == "" or query == "":
+    if st.button("Generate Response", key=map_query):
+        if endpoint_name == "" or map_query == "":
             st.error("Please enter a valid endpoint name and prompt!")
         else:
             with st.spinner("Wait for it..."):
-                generated_text = invoke_map_reduce(context, query, endpoint_name, os.getenv('region'))
+                generated_text = simple_map_reduce(context, map_query, reduce_query, endpoint_name, os.getenv('region'))
                 st.info(generated_text)
 
             st.success("Done!")