Merge branch 'develop' v0.1.11

CHANGELOG.md

@@ -6,6 +6,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.1.11] - 2023-11-07
+### Fixed
+- Error in Bedrock QnABotSettingQAPromptTemplate output - prompt does not terminate with '\n\nAssistant:` and generates error from Bedrock - #13
+
 ## [0.1.10] - 2023-10-27
 ### Fixed
 - Prompt bug: question not denoted by an XML tag, so LLM gets confused about what it's answering - PR #13

lambdas/bedrock-embeddings-and-llm/src/lambdahook.py

@@ -60,9 +60,15 @@ def get_generate_text(modelId, response):
         raise Exception("Unsupported provider: ", provider)
     return generated_text
 
+def replace_template_placeholders(prompt, event):
+    # history
+    history_array = json.loads(event["req"]["_userInfo"].get("chatMessageHistory","[]"))
+    history_str = '\n'.join(f"{key}: {value}" for item in history_array for key, value in item.items())
+    prompt = prompt.replace("{history}", history_str)
+    # TODO - replace additional prompt template placeholders - eg query, input, session attributes, user info
+    return prompt
 
-def format_prompt(modelId, prompt):
-    # TODO - replace prompt template placeholders - eg query, input, chatHistory, session attributes, user info  
+def format_prompt(modelId, prompt):  
     provider = modelId.split(".")[0]
     if provider == "anthropic":
         print("Model provider is Anthropic. Checking prompt format.")
@@ -75,10 +81,8 @@ def format_prompt(modelId, prompt):
     print(f"Prompt: {json.dumps(prompt)}")
     return prompt
 
-def get_llm_response(parameters, prompt):
+def get_llm_response(modelId, parameters, prompt):
     global client
-    modelId = parameters.pop("modelId", DEFAULT_MODEL_ID)
-    prompt = format_prompt(modelId, prompt)
     body = get_request_body(modelId, parameters, prompt)
     print("ModelId", modelId, "-  Body: ", body)
     if (client is None):
@@ -127,10 +131,13 @@ def lambda_handler(event, context):
     print("Received event: %s" % json.dumps(event)) 
     # args = {"Prefix:"<Prefix|None>", "Model_params":{"modelId":"anthropic.claude-instant-v1", "max_tokens":256}, "Prompt":"<prompt>"}
     args = get_args_from_lambdahook_args(event)
+    model_params = args.get("Model_params",{})
+    modelId = model_params.pop("modelId", DEFAULT_MODEL_ID)
     # prompt set from args, or from req.question if not specified in args.
     prompt = args.get("Prompt", event["req"]["question"])
-    model_params = args.get("Model_params",{})
-    llm_response = get_llm_response(model_params, prompt)
+    prompt = format_prompt(modelId, prompt)
+    prompt = replace_template_placeholders(prompt, event)
+    llm_response = get_llm_response(modelId, model_params, prompt)
     prefix = args.get("Prefix","LLM Answer:")
     event = format_response(event, llm_response, prefix)
     print("Returning response: %s" % json.dumps(event))

lambdas/bedrock-embeddings-and-llm/src/settings.py

@@ -3,7 +3,7 @@
 
 # Default prompt templates
 AMAZON_GENERATE_QUERY_PROMPT_TEMPLATE = """Given the following conversation and a follow up question, rephrase the follow up question to be a standalone question.<br>Chat History: <br>{history}<br>Follow up question: {input}<br>Standalone question:"""
-AMAZON_QA_PROMPT_TEMPLATE = """<br><br>Human: You are a friendly AI assistant. Answer the question in <question> tags only based on the provided reference passages. Here are reference passages in <references> tags:<br><references><br>{context}<br></references><br>If the references contain the information needed to respond, then write a confident response in under 50 words, quoting the relevant references. <br>Otherwise, if you can make an informed guess based on the reference passages, then write a less confident response in under 50 words, stating your assumptions.<br>Finally, if the references do not have any relevant information, then respond saying \\"Sorry, I don't know\\".<br><question><br>{query}<br></question><br>Assistant: According to the reference passages, in under 50 words:"""
+AMAZON_QA_PROMPT_TEMPLATE = """<br><br>Human: You are a friendly AI assistant. Answer the question in <question> tags only based on the provided reference passages. Here are reference passages in <references> tags:<br><references><br>{context}<br></references><br>If the references contain the information needed to respond, then write a confident response in under 50 words, quoting the relevant references. <br>Otherwise, if you can make an informed guess based on the reference passages, then write a less confident response in under 50 words, stating your assumptions.<br>Finally, if the references do not have any relevant information, then respond saying \\"Sorry, I don't know\\".<br><question><br>{query}<br></question><br><br>Assistant: According to the reference passages, in under 50 words:"""
 ANTHROPIC_GENERATE_QUERY_PROMPT_TEMPLATE = """<br><br>Human: Here is a chat history in <chatHistory> tags:<br><chatHistory><br>{history}<br></chatHistory><br>Human: And here is a follow up question or statement from the human in <followUpMessage> tags:<br><followUpMessage><br>{input}<br></followUpMessage><br>Human: Rephrase the follow up question or statement as a standalone question or statement that makes sense without reading the chat history.<br><br>Assistant: Here is the rephrased follow up question or statement:"""
 ANTHROPIC_QA_PROMPT_TEMPLATE = AMAZON_QA_PROMPT_TEMPLATE
 AI21_GENERATE_QUERY_PROMPT_TEMPATE = ANTHROPIC_GENERATE_QUERY_PROMPT_TEMPLATE

lambdas/bedrock-embeddings-and-llm/template.yml

@@ -1,5 +1,5 @@
 AWSTemplateFormatVersion: "2010-09-09"
-Description: QnABot on AWS LLM Plugin for Bedrock - v0.1.7
+Description: QnABot on AWS LLM Plugin for Bedrock - v0.1.11
 
 Parameters:
 
@@ -309,7 +309,7 @@ Resources:
       ServiceToken: !GetAtt OutputSettingsFunction.Arn
       EmbeddingsModelId: !Ref EmbeddingsModelId
       LLMModelId: !Ref LLMModelId
-      LastUpdate: '10/25/2023' 
+      LastUpdate: '11/07/2023' 
 
   TestBedrockModelFunction:
     Type: AWS::Lambda::Function