implement missing features in app.py, add retriever with summarization to scripts

app.py

@@ -3,29 +3,50 @@ import os
 import streamlit as st
 import tempfile
 
-from scripts import generate_metadata, ingest, MODEL_NAME
+from scripts import analyze_metadata, generate_metadata, ingest, MODEL_NAME
 
 
-st.title('DocVerifyRAG')
-st.write('Anomaly detection for BIM document metadata')
+st.title('# DocVerifyRAG')
+st.write('## Anomaly detection for BIM document metadata')
 
-uploaded_file = st.file_uploader("Choose a PDF file", type=["pdf","txt"])
+st.write('### Enter your file metadata in the following schema:')
 
-if uploaded_file is not None:
+user_input = st.text_input(
+    label='Filename, Description, Discipline',
+    value="", placeholder=str)
+
+if st.button('Submit'):
     try:
-        with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(uploaded_file.name)[1]) as tmp:
-            tmp.write(uploaded_file.read())
-            file_path = tmp.name
-            st.write(f'Created temporary file {file_path}')
-
-        docs = ingest(file_path)
-        st.write('## Querying Together.ai API')
-        metadata = generate_metadata(docs)
-        st.write(f'## Metadata Generated by {MODEL_NAME}')
-        st.write(metadata)
-
-        # Clean up the temporary file
-        os.remove(file_path)
-
-    except Exception as e:
-        st.error(f'Error: {e}')
+        filename, description, discipline = user_input.split(',')
+
+        st.write('## Analyzing with Vectara + together.ai')
+        analysis = analyze_metadata(filename, description, discipline)
+
+        st.write(analysis)
+
+        st.write('## Generate metadata?')
+        st.write('### Upload the file that corresponds to the submitted metadata')
+
+        uploaded_file = st.file_uploader("Choose a PDF file", type=["pdf","txt"])
+
+        if uploaded_file is not None:
+            try:
+                with tempfile.NamedTemporaryFile(delete=False, suffix=os.path.splitext(uploaded_file.name)[1]) as tmp:
+                    tmp.write(uploaded_file.read())
+                    file_path = tmp.name
+                    st.write(f'Created temporary file {file_path}')
+
+                docs = ingest(file_path)
+                st.write('## Querying Together.ai API')
+                metadata = generate_metadata(docs)
+                st.write(f'## Metadata Generated by {MODEL_NAME}')
+                st.write(metadata)
+
+                # Clean up the temporary file
+                os.remove(file_path)
+
+            except Exception as e:
+                st.error(f'Error: {e}')
+    except ValueError:
+        st.error('Please enter 3 comma separated values')
+

scripts.py

@@ -7,13 +7,54 @@ import sys
 from dotenv import load_dotenv
 from langchain_community.document_loaders import TextLoader
 from langchain_community.document_loaders import UnstructuredPDFLoader
-from langchain_community.embeddings.fake import FakeEmbeddings
+from langchain_community.embeddings import HuggingFaceEmbeddings
+from langchain_community.vectorstores import Vectara
+from langchain_core.output_parsers import StrOutputParser
+from langchain_core.prompts import ChatPromptTemplate
+from langchain_core.runnables import RunnablePassthrough
+from langchain.prompts import PromptTemplate
 from langchain_text_splitters import RecursiveCharacterTextSplitter
 
+
 load_dotenv()
 
 MODEL_NAME = "mistralai/Mixtral-8x7B-Instruct-v0.1"
 
+vectara_customer_id = os.environ['VECTARA_CUSTOMER_ID']
+vectara_corpus_id = os.environ['VECTARA_CORPUS_ID']
+vectara_api_key = os.environ['VECTARA_API_KEY']
+
+embeddings = HuggingFaceEmbeddings(model_name="intfloat/multilingual-e5-large")
+
+vectara = Vectara(vectara_customer_id=vectara_customer_id,
+                      vectara_corpus_id=vectara_corpus_id,
+                      vectara_api_key=vectara_api_key)
+
+
+summary_config = {"is_enabled": True, "max_results": 3, "response_lang": "eng"}
+retriever = vectara.as_retriever(
+    search_kwargs={"k": 3, "summary_config": summary_config}
+)
+
+template = """
+passage: You are a helpful assistant that understands BIM building documents.
+passage: You will analyze BIM document metadata composed of filename, description, and engineering discipline.
+passage: The metadata is written in German.
+passage: Filename: {filename}, Description: {description}, Engineering discipline: {discipline}.
+query: Does the filename match other filenames within the same discipline?
+query: Does the description match the engineering discipline?
+query: How different is the metadata to your curated information?
+query: Highligh any discrepancies and comment on wether or not the metadata is anomalous.
+"""
+
+prompt = PromptTemplate(template=template, input_variables=['filename', 'description', 'discipline'])
+
+
+def get_sources(documents):
+    return documents[:-1]
+
+def get_summary(documents):
+    return documents[-1].page_content
 
 def ingest(file_path):
     extension = os.path.splitext(file_path)[1].lower()
@@ -52,7 +93,7 @@ def generate_metadata(docs):
 
     You are a helpful assistant that understands BIM documents and engineering disciplines. Your answer should be in JSON format and only include the filename, a short description, and the engineering discipline the document belongs to, distinguishing between {[d.value for d in BimDiscipline]} based on the given document."
 
-    Analyze the provided document, which could be in either German or English. Extract the filename, its description, and infer the engineering discipline it belongs to. Document:
+    Analyze the provided document, which could be in either German or English. Extract the filename, its description, and infer the engineering discipline it belongs to. Respond in both English and German. Document:
     context="
     """     
     # plain text     
@@ -89,6 +130,11 @@ def generate_metadata(docs):
     return json.loads(chat_completion.choices[0].message.content)    
 
 
+def analyze_metadata(filename, description, discipline):
+    formatted_prompt = prompt.format(filename=filename, description=description, discipline=discipline)
+    return (retriever | get_summary).invoke(formatted_prompt)
+
+
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Generate metadata for a BIM document")
     parser.add_argument("document", metavar="FILEPATH", type=str,