Migrate Project to React

.gitignore

@@ -0,0 +1,28 @@
+# General
+.DS_Store
+__pycache__/
+
+# VS Code
+.vscode/*
+!.vscode/settings.json
+!.vscode/tasks.json
+!.vscode/launch.json
+!.vscode/extensions.json
+!.vscode/*.code-snippets
+.history/
+*.vsix
+
+# Dependency directories
+node_modules/
+
+# Database files
+requests.db
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/

README.md

@@ -1,5 +1,56 @@
-# InterviewAI
+A system that encourages writers to think deeply by asking them questions.
 
-With the advent of machine learning, it has become increasingly clear that AI is capable of completing tasks that were hitherto considered only possible by human minds. We are now pushing the boundaries of what AI can do with natural language processing (NLP), from summarizing pages of text to keeping up a conversation with a human. Our project aims to join those on the frontier of machine learning by creating an AI Interviewer. There are two main problems to address here: first, whether creating such an interviewer will be possible, and second, whether it will be any good.
+## Flourishing in this project looks like
 
-See our [project website](https://haramkoo.github.io/InterviewAI/) for more information.
+- Playful experimentation with user interfaces and question generation approaches
+- A shared place where we're running quantitative evaluations (perhaps quick-and-dirty) of different approaches
+- Eagerness to have writers try out this system
+- A draft of a paper coming together (perhaps starting with the formative study writing)
+- A presentation coming together, where we copy and paste examples of cool stuff (and broken stuff)
+
+
+
+## DAISI (Deep Artificial Intelligence System for Interviews)
+
+### About
+
+With the advent of machine learning, it has become increasingly clear that AI is capable of completing tasks that were hitherto considered only possible by human minds. We hope to join those on the frontier of machine learning by creating an AI Interviewer, temporarily named DAISI. There are two main problems to address here: first, whether creating such an interviewer will be possible, and second, whether it will be any good. We have managed thus far to create prototype models that are capable of asking questions, and we have created qualitative metrics to score the models' performance. We have also created some quantitative metrics based on those qualitative metrics, and hope to expand our quantitative metrics further with more research. Our next step as of the writing of this About section is to iteratively create more models that will hopefully perform better than our current prototypes.
+
+___
+
+### Project
+
+This is a research project; the focus is less on developing new software and more on refining pre-existing ones to suit our needs. Our research objective is first and foremost to see whether modern AI models can replicate the job of an interviewer to a satisfactory enough degree that we could feasibly use it for actual interviews. Should we succeed, there could be a use for the AI Interviewer, such as replacing journalists in dangerous locations like warzones. Moreover, the success of our AI Interviewer could aid in other research projects, such as AI to help writers, AI call center representatives, or any other position that involves asking questions or guiding conversations based on context.
+
+___
+
+### The Team
+
+#### Hyechan Jun
+
+Hyechan Jun is a Senior Computer Science (BCS) major with a double major in Writing. He plans to continue studying AI into graduate school, with the goal of obtaining a Ph.D.
+
+#### Ha-Ram Koo
+
+Ha-Ram Koo is a Senior Computer Science (BCS) major with a minor in Spanish. He is in charge of data collection and data wrangling procedures for this project.
+
+#### Advait Scaria
+
+Advait Scaria is a Senior Computer Science (BCS) major with a minor in Data Science. He is the main analyst and judge of the models' performance.
+
+___
+
+### Project Materials
+
+- [Project Website](https://haramkoo.github.io/InterviewAI/)
+- [Proposal](https://docs.google.com/document/d/1jYFH2y0SFvw5K9eVFQRNMO-UazR5w9YKnHaMpmKdFhk/edit?usp=sharing)
+- [Project Repo](https://github.com/haramkoo/InterviewAI)
+- [Research Paper/Report](https://drive.google.com/file/d/1tUyPhk9ePAeYYJh_T6nqyXttOsvmFExE/view?usp=sharing)
+- [Status Report Presentation](https://docs.google.com/presentation/d/1yWd4C6-0VyIv4WeSGReReajeSNEWJ7VndQVG7goXSJk/edit?usp=sharing)
+- [Web App](https://huggingface.co/spaces/calvininterview/interview-streamlit)
+
+<!-- Add link to final presentation -->
+
+### Other Links
+
+- [Calvin Computer Science Department](https://computing.calvin.edu/)

TODO.md

@@ -0,0 +1,57 @@
+# Issues
+
+- Need Error handling 
+
+## Prompts need improvement
+
+1. Fix ChatGPT prompt so that it is confined to text only provided by the user
+2. Find the prompt kca had written for question summaries
+
+Overall: try the prompt evaluation tool https://github.com/typpo/promptfoo
+
+We can test for hallucination by asserting that the response *doesn't* include some things that we know are not in the text.
+
+Don't bother asking about paragraphs that are too short (3 sentences or less?).
+
+Rate limiting on the server side.
+
+Only request reflections on parapgrahs right around the current one.
+
+- Put back in        "phrase_from_paragraph": "..."
+- Ensure that it actually matches a phrase from the parapgraph
+  - fuzzy match w/ regex?? (Ray)
+
+
+
+## Card display
+
+- Would be nice to show more than just the current paragraph (maybe the previous and next paragraph as well)
+- Response formatting: lists should show up as lists (Markdown?? Ask for a JSON response??)
+  - See the OpenAI "function calling" example for how to specify the response format
+  - Ask for the response in HTML? (Cheating...?)
+- Affordances
+  - Thumbs up and thumbs down buttons?? 
+    - Thumbs down regenerates? bust the cache somehow
+      - just log the thumbs-down?
+    - Thumbs up saves to a comment? log this somehow too
+
+
+## Two-way mapping
+
+- When you hover over a card, highlight the corresponding text in the document. (At least the corresponding paragraph.)
+- When you click on a paragraph, highlight the corresponding card. (Sophia has code that might work already.)
+
+## Controls
+
+- More preset prompts
+- Clearer way to edit the prompt?
+- Use FabricUI styles so it matches Word?
+- Maybe request multiple responses? Pick best of...?
+
+
+
+Alternative highlighting approach:
+
+
+Rewrite the following paragraph, put square brackets around the phrases that are relevant to "...".
+

interview-streamlit/README.md

@@ -1,12 +1,16 @@
----
-title: Interview Streamlit
-emoji: 📉
-colorFrom: pink
-colorTo: green
-sdk: streamlit
-sdk_version: 1.2.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces#reference
+# Interview AI Test Website
+
+This is an Interview AI testing stack designed by the [Senior Project 2021](https://haramkoo.github.io/InterviewAI/) team using the [Streamlit](https://streamlit.io/) framework. The production is meant to be live [here](https://huggingface.co/spaces/calvininterview/interview-streamlit), but it is currently down for unknown reasons.
+
+Run the following commands to run the front-end in your machine:
+
+*It is highly recommended that you run these commands in some sort of [Python virtual environment](https://www.youtube.com/watch?v=N5vscPTWKOk).*
+
+```
+pip install -r requirements.txt
+```
+```
+streamlit run app.py
+```
+
+The last command should launch the app in your favorite browser.

interview-streamlit/app.py

@@ -1,11 +1,6 @@
 import streamlit as st
-import torch
-from pandas import options
-from transformers import BartForConditionalGeneration, BartTokenizer
 
-# initialize model + tok variables
-model = None
-tok = None
+from backend import genQuestion
 
 # Examples for each models
 context_example = ''
@@ -17,50 +12,6 @@
     "I think things out well. When I speak, I speak with conviction. If I feel like it's something that best suits me and my person, I deal with it. I say it. I have no problem speaking out publicly about issues. But for personal things, and for things about personal selfishness, or wanting more money, I don't do that. Once I give my word, that's it. I don't go back to renegotiate. I don't renegotiate my contracts."
 ]
 
-
-# pass in Strings of model choice and input text for context
-@st.cache
-def genQuestion(model_choice, context, tag):
-    # global descriptions
-    if model_choice=="Base model":
-        model = BartForConditionalGeneration.from_pretrained("hyechanjun/interview-question-remake")
-        tok = BartTokenizer.from_pretrained("hyechanjun/interview-question-remake")
-    elif model_choice=="Lengthed model":
-        model = BartForConditionalGeneration.from_pretrained("hyechanjun/interview-length-tagged")
-        tok = BartTokenizer.from_pretrained("hyechanjun/interview-length-tagged")
-        if (tag == '1-10'):
-            context += ' <TEN>'
-        elif (tag == '11-20'):
-            context += ' <TWENTY>'
-        elif (tag == '21-30'):
-            context += ' <THIRTY>'
-        elif (tag == '31-40'):
-            context += ' <FORTY>'
-        elif (tag == '51-60'):
-            context += ' <FIFTY>'
-        elif (tag == '61-70'):
-            context += ' <SIXTY>'
-        elif (tag == '71-80'):
-            context += ' <SEVENTY>'
-        elif (tag == '81-90'):
-            context += ' <EIGHTY>'
-        elif (tag == '81-90'):
-            context += ' <NINETY>'
-        elif (tag == '91+'):
-            context += ' <HUNDRED>'
-    elif model_choice=="Reverse model":
-        model = BartForConditionalGeneration.from_pretrained("hyechanjun/reverse-interview-question")
-        tok = BartTokenizer.from_pretrained("hyechanjun/reverse-interview-question")
-
-    inputs = tok(context, return_tensors="pt")
-    output = model.generate(inputs["input_ids"], num_beams=4, max_length=64, min_length=9, num_return_sequences=4, diversity_penalty=1.0, num_beam_groups=4)
-    final_output = ''
-    for i in range(4):
-        final_output +=  [tok.decode(beam, skip_special_tokens=True, clean_up_tokenization_spaces=False) for beam in output][i] + "\n\n"
-
-    return final_output
-
-
 # Wide page layout
 st.set_page_config(layout="wide")
 
@@ -75,8 +26,8 @@ def genQuestion(model_choice, context, tag):
 col1, col2, col3 = st.columns(3)
 
 context_option = col2.selectbox(
-                'Feel free to choose one of our premade contexts',
-                ('Select one','Elon Musk', 'Fashion designer', 'Young entrepreneur', 'Michael Jordan')
+    'Feel free to choose one of our premade contexts',
+    ('Select one','Elon Musk', 'Fashion designer', 'Young entrepreneur', 'Michael Jordan')
 )
 
 if context_option == 'Select one':
@@ -91,10 +42,9 @@ def genQuestion(model_choice, context, tag):
     context_example = examples[3]
 
 option = col1.selectbox(
-                'Please select a model.',
-                ('Base model', 'Lengthed model', 'Reverse model'))
-
-
+    'Please select a model.',
+    ('Base model', 'Lengthed model', 'Reverse model')
+)
 
 if option == 'Base model':
     st.write("This is the re-fine-tuned base model for our interview AI. It returns strings terminating in a question mark (?).")
@@ -104,20 +54,18 @@ def genQuestion(model_choice, context, tag):
     st.write("This model asks a question that would have resulted in the context you provide (a.k.a. it traverses backward through the interview)")
 
 if option == 'Lengthed model':
-    context_length = col3.selectbox('Length of response',
-                                    ('1-10', '11-20', '21-30', '31-40', '41-50', '51-60', '61-70', '71-80', '81-90', '91+'))
-
+    context_length = col3.selectbox(
+        'Length of response',
+        ('1-10', '11-20', '21-30', '31-40', '41-50', '51-60', '61-70', '71-80', '81-90', '91+')
+    )
 
 # Input fields
-input = st.text_area('Context', value=context_example)                                    # user inputs context to construct a response (str)
-
+# user inputs context to construct a response (str)
+input = st.text_area('Context', value=context_example)
 
 if st.button('Submit') or st.session_state.button_sent:
     with st.spinner('Generating a response...'):
         output = genQuestion(option, input, context_length)
         print(output)
     st.session_state.button_sent = True
     st.text_area(label="Generated Responses:", value=output, height=200)
-
-
-

interview-streamlit/backend.py

@@ -0,0 +1,62 @@
+from transformers import BartForConditionalGeneration, BartTokenizer
+import streamlit as st
+
+@st.cache_data
+def genQuestion(model_choice, context, tag):
+    """Generate interview questions.
+
+    Args:
+        model_choice: specify which BART model to use.
+        context: input text for the model.
+        tag: specify the length of the generated question (only if model_choice is 'Lengthed model').
+
+    Returns:
+        final_output: the four decoded generated sequences, separated by two newlines.
+    """
+
+    model, tok = None, None
+
+    # Checks which BART model to use based on the value of model_choice
+    # Then loads a pre-trained BART model and tokenizer accordingly
+    if model_choice=="Base model":
+        model = BartForConditionalGeneration.from_pretrained("hyechanjun/interview-question-remake")
+        tok = BartTokenizer.from_pretrained("hyechanjun/interview-question-remake")
+    elif model_choice=="Lengthed model":
+        model = BartForConditionalGeneration.from_pretrained("hyechanjun/interview-length-tagged")
+        tok = BartTokenizer.from_pretrained("hyechanjun/interview-length-tagged")
+        if (tag == '1-10'):
+            context += ' <TEN>'
+        elif (tag == '11-20'):
+            context += ' <TWENTY>'
+        elif (tag == '21-30'):
+            context += ' <THIRTY>'
+        elif (tag == '31-40'):
+            context += ' <FORTY>'
+        elif (tag == '51-60'):
+            context += ' <FIFTY>'
+        elif (tag == '61-70'):
+            context += ' <SIXTY>'
+        elif (tag == '71-80'):
+            context += ' <SEVENTY>'
+        elif (tag == '81-90'):
+            context += ' <EIGHTY>'
+        elif (tag == '81-90'):
+            context += ' <NINETY>'
+        elif (tag == '91+'):
+            context += ' <HUNDRED>'
+    elif model_choice=="Reverse model":
+        model = BartForConditionalGeneration.from_pretrained("hyechanjun/reverse-interview-question")
+        tok = BartTokenizer.from_pretrained("hyechanjun/reverse-interview-question")
+
+    # Tokenize context parameter using the selected tokenizer
+    inputs = tok(context, return_tensors="pt")
+
+    # Generate text using the selected pre-trained BART model
+    output = model.generate(inputs["input_ids"], num_beams=4, max_length=64, min_length=9, num_return_sequences=4, diversity_penalty=1.0, num_beam_groups=4)
+
+    # Decode the generated text, then format it into a single string
+    final_output = ''
+    for i in range(4):
+        final_output +=  [tok.decode(beam, skip_special_tokens=True, clean_up_tokenization_spaces=False) for beam in output][i] + "\n\n"
+
+    return final_output