Added and configured a component to trim the architectures to remove some common junk referencing the vector retrieved documents.

config/architectures.json

@@ -16,7 +16,8 @@
       "steps": [
         {"class":  "InputRequestScreener"},
         {"class":  "RetrievalAugmentor", "params": {"vector_store": "01_all_products"}},
-        {"class":  "HFLlamaHttpRequestor", "params":  {"model":  "meta-llama/Llama-2-7b-chat-hf", "system_prompt": "You are a helpful domestic appliance advisor. Please answer the following customer question, answering only from the facts provided. Do not make things up, and say if you cannot answer.", "max_tokens": 2000}},
+        {"class":  "HFLlamaHttpRequestor", "params":  {"model":  "meta-llama/Llama-2-7b-chat-hf", "system_prompt": "You are a helpful domestic appliance advisor. Please answer the following customer question, answering only from the facts provided. Answer based on the background provided, do not make things up, and say if you cannot answer.", "max_tokens": 2000}},
+        {"class": "ResponseTrimmer", "params": {"regexes":  ["^.*information provided[0-9A-Za-z,]*? ", "^.*background[0-9A-Za-z,]*? "]}},
         {"class":  "OutputResponseScreener"}
       ],
       "img": "architecture_rag.jpg"

src/architectures.py

@@ -6,6 +6,7 @@ architecture components.
 import chromadb
 import json
 import os
+import regex as re
 import traceback
 
 from abc import ABC, abstractmethod
@@ -306,8 +307,12 @@ class RetrievalAugmentor(ArchitectureComponent):
         documents = results['documents'][0]  # Index 0 as we are always asking one question
 
         # Update the request to include the retrieved documents
-        new_query = f'QUESTION: {input_query}\n\n'
-        new_query += '\n'.join([f'FACT: {d}' for d in documents])
+        #new_query = f'QUESTION: {input_query}\n\n'
+        #new_query += '\n'.join([f'FACT: {d}' for d in documents])
+        new_query = '{"background": ['
+        new_query += ', '.join([f'"{d}"' for d in documents])
+        new_query += ']}\n\nQUESTION: '
+        new_query += input_query
 
         # Put the request back into the architecture request
         request.request = new_query
@@ -356,3 +361,27 @@ class HFLlamaHttpRequestor(ArchitectureComponent):
             raise ValueError(f'No model {self.model} configured in the environment')
         response = llm(request.request, system_prompt=self.system_prompt, max_new_tokens=self.max_tokens, temperature=self.temperature)
         request.response = response
+
+
+class ResponseTrimmer(ArchitectureComponent):
+    """
+    A concrete pipeline component which trims the response based on a regex match,
+    then uppercases the first character of what is left.
+    """
+    description = "Trims the response based on a regex"
+
+    def __init__(self, regexes: List[str]):
+        quoted_regexes = [f'"{r}"' for r in regexes]
+        self.regex_display = f"[{', '.join(quoted_regexes)}]"
+        self.regexes = [re.compile(r, re.IGNORECASE) for r in regexes]
+
+    def process_request(self, request: ArchitectureRequest):
+        new_response = request.response
+        for regex in self.regexes:
+            new_response = regex.sub('', new_response)
+        new_response = new_response[:1].upper() + new_response[1:]
+        request.response = new_response
+
+    def config_description(self) -> str:
+        return f"Regexes: {self.regex_display}"
+