Update app.py

app.py

@@ -59,11 +59,26 @@ def save_key(api_key):
     return api_key
 
 
-def query_pinecone(query, top_k, model, index):
+def query_pinecone(query, top_k, model, index, year, quarter, ticker, threshold=0.5):
     # generate embeddings for the query
     xq = model.encode([query]).tolist()
     # search pinecone index for context passage with the answer
-    xc = index.query(xq, top_k=top_k, include_metadata=True)
+    xc = index.query(
+        xq,
+        top_k=top_k,
+        filter={
+            "Year": int(year),
+            "Quarter": {"$eq": quarter},
+            "Ticker": {"$eq": ticker},
+        },
+        include_metadata=True,
+    )
+    # filter the context passages based on the score threshold
+    filtered_matches = []
+    for match in xc["matches"]:
+        if match["score"] >= threshold:
+            filtered_matches.append(match)
+    xc["matches"] = filtered_matches
     return xc
 
 
@@ -96,13 +111,10 @@ def text_lookup(data, sentence_ids):
     return context
 
 
-def gpt3_summary(result, query):
-    prompt = f"""Answer the question based on the following information:
-                {result}
-                Question: {query} """
+def gpt3_summary(text):
     response = openai.Completion.create(
-        model="text-ada-001",
-        prompt=  prompt,
+        model="text-davinci-003",
+        prompt=text + "\n\nTl;dr",
         temperature=0.1,
         max_tokens=512,
         top_p=1.0,
@@ -126,77 +138,115 @@ def gpt3_qa(query, answer):
     return response.choices[0].text
 
 
-st.title("Abstractive Question Answering - APPL")
+st.title("Abstractive Question Answering")
+
+st.write(
+    "The app uses the quarterly earnings call transcripts for 10 companies (Apple, AMD, Amazon, Cisco, Google, Microsoft, Nvidia, ASML, Intel, Micron) for the years 2016 to 2020."
+)
 
 query_text = st.text_input("Input Query", value="Who is the CEO of Apple?")
 
-num_results = int(st.number_input("Number of Results to query", 1, 5, value=2))
+years_choice = ["2016", "2017", "2018", "2019", "2020"]
+
+year = st.selectbox("Year", years_choice)
+
+quarter = st.selectbox("Quarter", ["Q1", "Q2", "Q3", "Q4"])
+
+ticker_choice = [
+    "AAPL",
+    "CSCO",
+    "MSFT",
+    "ASML",
+    "NVDA",
+    "GOOGL",
+    "MU",
+    "INTC",
+    "AMZN",
+    "AMD",
+]
+
+ticker = st.selectbox("Company", ticker_choice)
+
+num_results = int(st.number_input("Number of Results to query", 1, 5, value=3))
 
 
 # Choose encoder model
 
-encoder_models_choice = ["MPNET", "SGPT"]
+encoder_models_choice = ["SGPT", "MPNET"]
 
 encoder_model = st.selectbox("Select Encoder Model", encoder_models_choice)
 
 
 # Choose decoder model
 
-decoder_models_choice = ["GPT3 (QA_davinci)", "GPT3 (text_davinci)", "T5", "FLAN-T5"]
+decoder_models_choice = ["FLAN-T5", "T5", "GPT3 (QA_davinci)", "GPT3 (summary_davinci)"]
 
 decoder_model = st.selectbox("Select Decoder Model", decoder_models_choice)
 
 
 if encoder_model == "MPNET":
     # Connect to pinecone environment
-    pinecone.init(
-        api_key="ea9fd320-6f8a-4edd-bf41-9e972b95cbf9", environment="us-east1-gcp"
-    )
+    pinecone.init(api_key=st.secrets["pinecone_mpnet"], environment="us-east1-gcp")
     pinecone_index_name = "week2-all-mpnet-base"
     pinecone_index = pinecone.Index(pinecone_index_name)
     retriever_model = get_mpnet_embedding_model()
 
 elif encoder_model == "SGPT":
     # Connect to pinecone environment
-    pinecone.init(
-        api_key="0d8215d7-4ad5-4c76-8c45-4a40c0f6a1b7", environment="us-east1-gcp"
-    )
+    pinecone.init(api_key=st.secrets["pinecone_sgpt"], environment="us-east1-gcp")
     pinecone_index_name = "week2-sgpt-125m"
     pinecone_index = pinecone.Index(pinecone_index_name)
     retriever_model = get_sgpt_embedding_model()
 
 
-query_results = query_pinecone(query_text, num_results, retriever_model, pinecone_index)
+window = int(st.number_input("Sentence Window Size", 0, 3, value=0))
 
-window = int(st.number_input("Sentence Window Size", 1, 3, value=1))
+threshold = float(
+    st.number_input(
+        label="Similarity Score Threshold", step=0.05, format="%.2f", value=0.55
+    )
+)
 
 data = get_data()
 
-# context_list = format_query(query_results)
-context_list = sentence_id_combine(data, query_results, lag=window)
+query_results = query_pinecone(
+    query_text,
+    num_results,
+    retriever_model,
+    pinecone_index,
+    year,
+    quarter,
+    ticker,
+    threshold,
+)
+
+if threshold <= 0.60:
+    context_list = sentence_id_combine(data, query_results, lag=window)
+else:
+    context_list = format_query(query_results)
 
 
 st.subheader("Answer:")
 
 
-if decoder_model == "GPT3 (text_davinci)":
+if decoder_model == "GPT3 (summary_davinci)":
     openai_key = st.text_input(
         "Enter OpenAI key",
-        value="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",
+        value=st.secrets["openai_key"],
         type="password",
     )
     api_key = save_key(openai_key)
     openai.api_key = api_key
-    #output_text = []
-    #for context_text in context_list:
-    #    output_text.append(gpt3_summary(context_text))
-    #generated_text = ". ".join(output_text)
-    st.write(gpt3_summary(context_list,query_text))
+    output_text = []
+    for context_text in context_list:
+        output_text.append(gpt3_summary(context_text))
+    generated_text = ". ".join(output_text)
+    st.write(gpt3_summary(generated_text))
 
 elif decoder_model == "GPT3 (QA_davinci)":
     openai_key = st.text_input(
         "Enter OpenAI key",
-        value="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",
+        value=st.secrets["openai_key"],
         type="password",
     )
     api_key = save_key(openai_key)
@@ -226,8 +276,6 @@ elif decoder_model == "FLAN-T5":
 show_retrieved_text = st.checkbox("Show Retrieved Text", value=False)
 
 if show_retrieved_text:
-
     st.subheader("Retrieved Text:")
-
     for context_text in context_list:
-        st.markdown(f"- {context_text}")
+        st.markdown(f"- {context_text}")