initial commit

.github/workflows/encode-faq-js.yml

@@ -0,0 +1,60 @@
+name: Generate TensorFlow.js Encodings
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "* 0 * * *"
+env:
+  FILEPATH: ./encode-faq-js/
+  FILENAME: stampy-encodings.json
+  PROJECT_NUM: 507517956350
+  PROJECT_ID: stampy-nlp
+  BUCKET_ID: stampy-nlp-resources
+  ID_POOL: stampy-pool
+  ID_PROVIDER: stampy-provider
+jobs:
+  encode-faq-js:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: 'read'
+      id-token: 'write'
+    steps:
+    - name: Clone Repository Code
+      id: 'checkout'
+      uses: 'actions/checkout@v3'
+      with:
+        persist-credentials: false # otherwise, the token used is the GITHUB_TOKEN, instead of your personal access token.
+        fetch-depth: 0 # otherwise, there would be errors pushing refs to the destination repository.
+
+    - name: Generate FAQ Encodings for TensorFlow.js
+      id: generate
+      run: |
+        cd ${{ env.FILEPATH }}
+        npm install node-fetch @tensorflow/tfjs-node@4.1.0 @tensorflow-models/universal-sentence-encoder
+        node encode-faq.mjs ${{ secrets.CODA_TOKEN }}
+
+    - name: 'Authenticate to Google Cloud'
+      id: 'auth'
+      uses: 'google-github-actions/auth@v1'
+      with:
+        credentials_json: '${{ secrets.GCP_CREDENTIALS }}'
+        # workload_identity_provider: 'projects/${{ env.PROJECT_NUM }}/locations/global/workloadIdentityPools/${{ env.ID_POOL }}/providers/${{ env.ID_PROVIDER }}'
+        # service_account: 'service@${{ env.PROJECT_ID }}.iam.gserviceaccount.com'
+        # token_format: 'access_token'
+
+    - name: 'Setup Cloud SDK'
+      uses: 'google-github-actions/setup-gcloud@v1'
+      with:
+        version: '>= 400.0.0'
+
+    - name: 'Debugging Info with Cloud CLI'
+      run: |
+        gcloud info
+        gcloud auth list
+        gcloud config list
+          
+    - name: 'Upload TFJS Encodings JSON to Cloud Bucket'
+      id: 'upload-file'
+      uses: 'google-github-actions/upload-cloud-storage@v1'
+      with:
+        path: '${{ env.FILEPATH }}${{ env.FILENAME }}'
+        destination: '${{ env.BUCKET_ID }}'

.github/workflows/encode-faq-py.yml

@@ -0,0 +1,55 @@
+name: Sync Coda to Pinecone Encodings
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "* 0 * * *"
+env:
+  FILEPATH: ./encode-faq-py/
+  FILENAME: stampy-duplicates.json
+  PROJECT_ID: stampy-nlp
+  BUCKET_ID: stampy-nlp-resources
+  CODA_TOKEN: ${{ secrets.CODA_TOKEN }}
+jobs:
+  encode-faq-py:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: 'read'
+      id-token: 'write'
+    steps:
+    - name: Clone Repository Code
+      id: 'checkout'
+      uses: 'actions/checkout@v3'
+      with:
+        persist-credentials: false # otherwise, the token used is the GITHUB_TOKEN, instead of your personal access token.
+        fetch-depth: 0 # otherwise, there would be errors pushing refs to the destination repository.
+
+    - name: 'Authenticate to Google Cloud'
+      id: 'auth'
+      uses: 'google-github-actions/auth@v1'
+      with:
+        credentials_json: '${{ secrets.GCP_CREDENTIALS }}'
+
+    - name: 'Setup Cloud SDK'
+      uses: 'google-github-actions/setup-gcloud@v1'
+      with:
+        version: '>= 400.0.0'
+
+    - name: 'Debugging Info with Cloud CLI'
+      run: |
+        gcloud info
+        gcloud auth list
+        gcloud config list
+          
+    - name: Generate FAQ Encodings and Upload to Pinecone Vector DB
+      id: generate
+      run: |
+        cd ${{ env.FILEPATH }}
+        pip install -q -r requirements.txt
+        python3 encode-faq-titles.py
+
+    - name: 'Upload Duplicates JSON to Cloud Bucket'
+      id: 'upload-file'
+      uses: 'google-github-actions/upload-cloud-storage@v1'
+      with:
+        path: '${{ env.FILEPATH }}${{ env.FILENAME }}'
+        destination: '${{ env.BUCKET_ID }}'

.gitignore

@@ -0,0 +1,12 @@
+credentials.json
+stampy-*.json
+models
+node_modules
+*.bin
+*.pyc
+__pycache__
+node_modules
+.DS_Store
+.git
+secrets.sh
+venv

README.md

@@ -0,0 +1,160 @@
+![Stampy!](https://github.com/StampyAI/StampyAIAssets/blob/main/profile/stampy-profile-228.png?raw=true)
+
+Stampy NLP performs semantic search and other NLP microservices for [aisafety.info](https://aisafety.info) and [stampy.ai](https://stampy.ai), a database of questions and answers about AGI safety. Contributions will be welcome (once I get the messy things cleaned up), and the code is released under the MIT License.
+
+The demo url is [nlp.stampy.ai](https://nlp.stampy.ai/) or direct link to [stampy-nlp-t6p37v2uia-uw.a.run.app](https://stampy-nlp-t6p37v2uia-uw.a.run.app/). If you're interested in learning more about Natural Language Processing (NLP) and Transformers, the [HuggingFace course](https://huggingface.co/course/chapter1/2?fw=pt) provides an excellent introduction.
+
+# Main Services Overview
+
+![Stampy NLP Overview](images/stampy-nlp.svg)
+
+Our NLP services offer 4 features which depend on 2 key components:
+
+1. Three NLP models from HuggingFace, specifically [SentenceTransformers](https://www.sbert.net/), provide pre-trained models optimized for different types of semantic searches by generating sentence embeddings -- 768-dimension vectors, numerical representations that capture the meaning of the text. Think of it as an 768 element array of floats. In general, we use Python + PyTorch since that gives us the most flexibility to use a variety of models by default.
+
+- Retriever model ([multi-qa-mpnet](https://huggingface.co/sentence-transformers/multi-qa-mpnet-base-cos-v1)) for identifying paraphrased questions.
+- [allenai-specter](https://huggingface.co/sentence-transformers/allenai-specter) for searching titles & abstracts of scientific publications.
+- Reader model ([electra-base-squad2](https://huggingface.co/deepset/electra-base-squad2)) finds the start & end index of the answer given a question and a context paragraph containing the answer.
+
+2. [Pinecone](https://www.pinecone.io/) is a fully managed, high-performance database for vector search applications. Each data element contains the 768-dimension vector, a unique id (i.e. Coda id for our FAQ) and some metadata (original text, url, other relevant information).
+
+## 1. [Semantic Search for Similar FAQs](https://nlp.stampy.ai/)
+
+Encodes a given `query` string, sends the vector embedding to search pinecone for nearest entries in `faq-titles` namespace, then returns payload as json sorted by score between 0 and 1 indicating the similarity of match.
+
+Sample API usage:
+
+```text
+https://nlp.stampy.ai/api/search?query=What+AI+safety%3F`
+```
+
+- `query` (required) is the sentence or phrase to be encoded then have nearest entries returned.
+- `top` (optional) indicates the number of entries returned. If the value is not specified, the default is to return the 10 nearest entries.
+- `showLive=0` (optional) will return only entries with `status` that are NOT "Live on site". The default is `showLive=1` to return only entries that with `status` that are "Live on site".
+- `status=all` (optional) returns all entries including those that have not yet been canonically answered. Specify multiple values for `status` to return matching more than one value.
+
+Additional usages:
+
+`showLive=1` returns entries where `status == "Live on site"`
+
+```bash
+https://stampy-nlp-t6p37v2uia-uw.a.run.app/api/search?query=What+AI+safety%3F&showLive=1
+```
+
+`showLive=0` returns entries where `status != "Live on site"`
+
+```bash
+https://stampy-nlp-t6p37v2uia-uw.a.run.app/api/search?query=What+AI+safety%3F&showLive=0  
+```
+
+`status=all` returns all questions regardless of status
+
+```bash
+https://stampy-nlp-t6p37v2uia-uw.a.run.app/api/search?query=Something+random%3F&status=all  
+```
+
+`status=value` returns entries with status matching whatever value is specified. Multiple values may be listed separately. The example below returns entries with `status == "Not started"` and also `status == "In progress"`
+
+```bash
+https://stampy-nlp-t6p37v2uia-uw.a.run.app/api/search?query=Something+random%3F&status=Not%20started&status=In%20progress 
+```
+
+## 2. [Duplicates Report](https://nlp.stampy.ai/duplicates)
+
+Display a table with top pairs of most similar questions in Coda based on the last time `paraphrase_mining` was called.
+
+## 3. [Literature Abstracts](https://nlp.stampy.ai/literature)
+
+Encodes a given query string, sends the vector embedding to search pinecone for nearest entry in `paper-abstracts` namespace, then returns payload as json sorted by score between 0 and 1 indicating the similarity of match. In an effort to minimize the number of huge models in our app container, this service still uses the external HuggingFace API so it's still a bit slow.
+
+Sample API usage:
+
+```text
+https://nlp.stampy.ai/api/literature?query=What+AI+safety%3F
+```
+
+## 4. [Extract QA](https://nlp.stampy.ai/extract)
+
+Encodes a given query string then sends the vector embedding to search pinecone for the 10 nearest entries in `extracted-chunks` namespace. For each entry, run the HuggingFace pipeline task to extract the answer from each content then returns payload as json sorted by score between 0 and 1 indicating the confidence of the answer matches the query question. Since this runs +10 inferences, this can be rather slow.
+
+Sample API usage:
+
+```text
+https://nlp.stampy.ai/api/extract?query=What+AI+safety%3F
+```
+
+# Setup Environment
+
+Obtain free API tokens for Coda, Pincecone, Hugging Face and set the environment variables below with their values.
+
+```text
+CODA_TOKEN=
+PINECONE_API_KEY=
+HUGGINGFACE_API_KEY=
+```
+
+## Install [Git Large File Storage](https://git-lfs.com)
+
+After clone this repo, the large Transformer models must be downloaded separately using `git lfs`.
+
+```bash
+curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
+sudo apt-get install git-lfs
+git lfs install
+```
+
+## Install Google [Cloud SDK](https://cloud.google.com/sdk/docs/install)
+
+Google Cloud SDK will be needed if you plan to upload generated files to Cloud bucket.
+
+```bash
+echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] https://packages.cloud.google.com/apt cloud-sdk main" | sudo tee -a /etc/apt/sources.list.d/google-cloud-sdk.list
+curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key --keyring /usr/share/keyrings/cloud.google.gpg add -
+sudo apt-get update && sudo apt-get install google-cloud-cli
+gcloud init
+gcloud auth login --no-launch-browser
+```
+
+# Files Overview
+
+`main.py` is the main Flask server that responds to API endpoints and also displays web pages demonstrating the use of API calls.
+
+Files named `encode_*.py` use the relevant model to encode the associated text then stores the results in a specific pinecone namespace that can be retrieved by the Flask server.
+
+- `encode_faq_titles.py` encodes all the question titles from the Coda, then stores into pinecone `faq-titles` namespace for later use in Semantic Search for Similar FAQs on Stampy. Also uses the SentenceTransformer utility `paraphrase_mining` to generate a JSON file with top 100 pairs of most similar questions in Coda.
+
+- `encode_extracted_chunks.py` takes alignment dataset (currently all ~1000 arxiv papers and ~2000 Alignment Forum blog posts), splits them into ~1000 word chunks split on complete sentences, encodes them, then stores into pinecone `extracted-chunks` namespace for later use in Extract QA.
+
+- `encode_lit_abstracts.py` takes all arxiv papers from alignment dataset, encodes the title + abstracts, then stored into pinecone `lit-abstracts` namespace for later use in Paper Search.
+
+# Discussion + Needs
+
+## 1. Automate Deployment
+
+Right now all this is deployed on Google Cloud Run, which allows all the dependencies to be containerized, supposedly starts up instantaneously and should scale to meet increased demand. Each of the HuggingFace NLP models are ~400 MB and we use 3 separate ones. Originally, to keep the app size small, I used the external Hugging Face API but it was very slow. To speed up inference, the transformer models are now packaged as part of the Cloud Run container, but that also means the app requires at least 4GB of memory to run as one "monolith." I believe this approach has addressed the latency issue for now. There are some recommendations to split the 3 models' functions into separate microservices which can run and scale independently, which is something I was planning to try next. Thoughts on other options?
+
+## 2. Keep Code Databases Synchronized
+
+Currently, a script is manually run to synchronize the list of questions in Coda to pinecone. Consider adding `/api/create` and `/api/delete` endpoints which take a Coda document id, synchronize data for the given entry in pinecone.
+
+```text
+https://nlp.stampy.ai/api/create?id=i-7dc3df6cd1e28f067c354042b0276673117624195a0adae01f83e34fbd2be652
+https://nlp.stampy.ai/api/delete?id=...
+```
+
+## 3. Keep Alignment Dataset Up-to-Date
+
+At the moment pinecone is free up to 1 million vectors, of which we are using ~40,000 vectors. It seems very responsive but there may be a price increase of we index the entire alignment literature dataset. It'd been nice to have someone who can do a cleaner parse of the [alignment dataset](https://github.com/moirage/alignment-research-dataset) and also Stampy's FAQ to extract the plain text from blogs, pdfs and other documents and keep alignment-related literature up-to-date with newly published arxiv papers and blog-posts. Also consider adding generated summaries of blog posts and treat like paper abstracts to expand paper search space.
+
+## 4. Clean Up Code
+
+Before all the above... The messy code needs to be cleaned up & refactored before getting others to contribute. I'm always open to suggestions in general!
+
+# Deploy to Google [Cloud Run](https://cloud.google.com/sdk/gcloud/reference/beta/run/deploy)
+
+```zsh
+chmod +x *.sh
+./setup.sh
+python3 main.py
+./deploy.sh
+```

api/.dockerignore

@@ -0,0 +1,3 @@
+.DS_Store
+.git
+env

api/Dockerfile

@@ -0,0 +1,20 @@
+FROM python:3.8
+
+# Allow statements and log messages to immediately appear in the Knative logs
+ENV PYTHONUNBUFFERED True
+
+# set a directory for the app
+WORKDIR /app
+
+# copy all the files to the container
+COPY . .
+
+# install dependencies
+RUN pip install --no-cache-dir -r requirements.txt
+
+# tell the port number the container should expose
+EXPOSE 8080
+ENV PORT 8080
+
+# Use gunicorn as the entrypoint
+CMD exec gunicorn --bind :$PORT --workers 1 --threads 1 --timeout 0 main:app

api/Procfile

@@ -0,0 +1 @@
+web: gunicorn --bind :$PORT --workers 1 --threads 1 --timeout 0 main:app

api/deploy.sh

@@ -0,0 +1,32 @@
+# echo "Activating Dev Env"
+# virtualenv -p /usr/local/bin/python3.8 venv
+# source venv/bin/activate
+
+# export CODA_TOKEN=
+# export PINECONE_API_KEY=
+# export HUGGINGFACE_API_KEY=
+
+echo "Exporting GCLOUD_PROJECT, CLOUD_RUN_SERVICE and LOCATION"
+export GCLOUD_PROJECT=stampy-nlp
+export GCLOUD_BUCKET=stampy-nlp-resources
+export CLOUD_RUN_SERVICE=stampy-nlp
+export LOCATION=us-west1
+
+echo "Deploying to Google Cloud Run"
+
+# gcloud builds submit --tag gcr.io/$GOOGLE_PROJECT_ID/$CLOUD_RUN_SERVICE \
+#   --project=$GOOGLE_PROJECT_ID
+
+# gcloud run deploy $CLOUD_RUN_SERVICE \
+#   --image gcr.io/$GOOGLE_PROJECT_ID/$CLOUD_RUN_SERVICE \
+#   --add-cloudsql-instances $INSTANCE_CONNECTION_NAME \
+#   --update-env-vars INSTANCE_CONNECTION_NAME=$INSTANCE_CONNECTION_NAME,DB_PASS=$DB_PASS,DB_USER=$DB_USER,DB_NAME=$DB_NAME \
+#   --platform managed \
+#   --region us-central1 \
+#   --allow-unauthenticated \
+#   --project=$GOOGLE_PROJECT_ID
+
+gcloud beta run deploy $CLOUD_RUN_SERVICE --source . \
+--min-instances=1 --memory 4G --cpu=2 --platform managed --no-traffic --tag=test \
+--service-account=service@stampy-nlp.iam.gserviceaccount.com \
+--update-secrets=PINECONE_API_KEY=PINECONE_API_KEY:latest,HUGGINGFACE_API_KEY=HUGGINGFACE_API_KEY:latest,CODA_TOKEN=CODA_TOKEN:latest