lambda for RAG

backend/deploy_lambda_retrieval_and_generation.sh

@@ -0,0 +1,16 @@
+#!/bin/sh
+rm -rf package
+rm package.zip
+mkdir -p package
+pip install -r requirements-lambda.txt --force-reinstall --target ./package --only-binary=:all: --platform manylinux1_x86_64 --upgrade
+pip install pydantic --force-reinstall --target ./package  --only-binary=:all: --platform manylinux1_x86_64 --upgrade
+pip install pydantic_core --force-reinstall --target ./package  --only-binary=:all: --platform manylinux1_x86_64 --upgrade
+cd package q
+zip -r ../package.zip .
+cd ..
+zip package.zip lambda_function_retrieval_and_generation.py
+zip package.zip config.py
+zip -r package.zip src
+aws --profile $AWS_PROFILE_NAME --region eu-central-1 lambda update-function-code \
+    --function-name  electionsAIRetrievalAndGeneration \
+    --zip-file fileb://package.zip

backend/deploy_lambda.sh → backend/deploy_lambda_retrieval_only.sh

@@ -6,7 +6,7 @@ pip install -r requirements-lambda.txt --force-reinstall --target ./package
 cd package 
 zip -r ../package.zip .
 cd ..
-zip package.zip lambda_function.py
+zip package.zip lambda_function_retrieval_only.py
 zip package.zip config.py
 zip -r package.zip src
 aws --profile $AWS_PROFILE_NAME --region eu-central-1 lambda update-function-code \

backend/lambda_function_retrieval_and_generation.py

@@ -0,0 +1,37 @@
+from src.bedrock import retrieve_bedrock_items, re_reference, BedrockRetrievedItem
+from src.llm import decisions_query
+from config import BEDROCK_KNOWLEDGE_BASE_ID
+import json
+from typing import Optional
+
+def lambda_handler(event: dict, context: Optional[dict] = None):
+
+    query = event["query"]
+
+    if len(query) < 10:
+        return {"statusCode": 400, "body": json.dumps("Vraag niet lang genoeg.")}
+    if len(query) > 500:
+        return {"statusCode": 400, "body": json.dumps("Vraag te lang.")}
+
+    retrieved_items = retrieve_bedrock_items(BEDROCK_KNOWLEDGE_BASE_ID, query, 10)
+    llm_response = decisions_query(query, retrieved_items)
+    re_referenced_llm_response, used_decisions = re_reference(llm_response, retrieved_items)
+
+    used_decisions_dicts = []
+    for decision in used_decisions:
+        used_decisions_dicts.append({
+            "text": decision.text,
+            "decision_url": decision.decision_url,
+            "title": decision.title,
+            "meeting_date": decision.meeting_date,
+            "score": decision.score
+        })
+
+    return {"statusCode": 200, "body": json.dumps({"response":re_referenced_llm_response,"decisions":used_decisions_dicts})}
+
+if __name__ == "__main__":
+    response = lambda_handler({
+            "query": "wat heeft de regering gedaan om de economie te stimuleren?"
+        }
+    )
+    print(response)

backend/lambda_function.py → backend/lambda_function_retrieval_only.py

@@ -9,7 +9,7 @@ def lambda_handler(event: dict, context: Optional[dict] = None):
     if len(event["query"]) < 10:
         return {"statusCode": 400, "body": json.dumps("Vraag niet lang genoeg.")}
 
-    if len(event["query"]) > 1000:
+    if len(event["query"]) > 500:
         return {"statusCode": 400, "body": json.dumps("Vraag te lang.")}
 
     response = retrieve_bedrock_items(

backend/requirements-lambda.txt

@@ -1,2 +1,3 @@
 boto3
-python-dateutil
+python-dateutil
+openai

backend/src/llm.py

@@ -1,16 +1,13 @@
 from typing import Optional
 import openai
-import tiktoken
 from config import OPENAI_API_KEY, OPENAI_MODEL_NAME
-from src.vector_store import VectorDBItem
 from src.bedrock import BedrockRetrievedItem
 # TODO: Bedrock items and vector db items need to combined into a single type
 
 client = openai.OpenAI(
     api_key=OPENAI_API_KEY
 )
 
-encoding = tiktoken.encoding_for_model(OPENAI_MODEL_NAME)
 
 def get_response(prompt:str, model_overide=None)->str:
     api_response = client.chat.completions.create(
@@ -26,9 +23,15 @@ def get_response(prompt:str, model_overide=None)->str:
     return api_response.choices[0].message.content
 
 def token_count(text:str)->int:
+    # tiktoken is not playing nicely on aws lambda so we need to do this 
+    # workaround to avoid importing it
+    import tiktoken
+    encoding = tiktoken.encoding_for_model(OPENAI_MODEL_NAME)
     return len(encoding.encode(text))
 
-def generate_prompt_for_decision_consult(topic_text:str, matching_decisions:list[VectorDBItem], max_tokens: int)->str:
+def generate_prompt_for_decision_consult(topic_text:str, matching_decisions, max_tokens: int)->str:
+    # we can't put a type on the decisions here because it requires importing from the vector db module
+    # and this is giving a problem when building the lambda... when there time - fix this
     flat_decisions = [f"{index} - {item.text}" for index, item in enumerate(matching_decisions)]
 
     prompt_instruction = """
@@ -92,7 +95,6 @@ def decisions_query(query:str, decisions:list[BedrockRetrievedItem])->str:
                 vraag:{query}
                 Jou antwoord:
     """
-    print(f"full prompt contains: {token_count(prompt)} tokens")
     return get_response(prompt)