Prepare for public source code release

.github/workflows/ci.yaml

@@ -25,6 +25,12 @@ jobs:
         run: npm run build --if-present
       - name: "Test"
         run: npm run test --if-present
+      - name: "Configure for deployment"
+        if: github.ref == 'refs/heads/master'
+        run: |
+          echo "$PROD_ENV" > lib/env.ts
+        env:
+          PROD_ENV: ${{ secrets.PROD_ENV }}
       - name: "Deploy"
         if: github.ref == 'refs/heads/master'
         run: npm run deploy

.gitignore

@@ -1,4 +1,5 @@
 /node_modules
 /.airtableblocksrc.json
 /build
-.tmp
+.tmp
+env.ts

README.md

@@ -1,6 +1,16 @@
 # ai-evaluator-extension ![deployment automatic](https://img.shields.io/badge/deployment-automatic-success)
 
-Airtable extension for evaluating application to our courses using AI, specifically large language models (LLMS).
+[Airtable extension](https://airtable.com/developers/extensions/guides/getting-started) for helping us evaluate applications to our courses using large language models (LLMs).
+
+![Screenshot of the Airtable extension](./screenshot.png)
+
+## Context
+
+We previously manually evaluated each applicant for our educational courses on a set of objective criteria. We would then use these scores plus some additional subjective judgement to come to an application decision. This is time-intensive, and it's difficult to align all the humans to give the same responses for the same person.
+
+In a [previous pilot project](https://github.com/bluedotimpact/ai-application-evaluations-pilot) we concluded that LLMs, while not perfect, could help us automate the first scoring part of our application process.
+
+This repository holds code for an Airtable extension that we can run inside our applications base. We set the relevant inputs (e.g. answers to application questions), the decisioning criteria, and let it evaluate applicants.
 
 ## Developer setup
 

lib/env.template.ts

@@ -0,0 +1,15 @@
+// For local development, you should usually edit env.ts (and not env.template.ts)
+// env.template.ts is a template to set up semi-valid env variables for new projects
+
+// This is very jank
+// We're building secrets in to the source code accessible in the browser. This is because Airtable extensions don't have any sensible secret management mechanism.
+// We're currently accepting this risk given:
+// - only BlueDot Impact employees have access to the deployed extension
+// - we have limits on API usage set for both platforms
+// - the risk is people using our API credits / getting free access to APIs, which is not as critical (as e.g. applicant's personal data)
+export const env = {
+    // Get from https://platform.openai.com/settings/profile?tab=api-keys
+    'OPENAI_API_KEY': 'sk-...',
+    // Get from https://console.anthropic.com/settings/keys
+    'ANTHROPIC_API_KEY': 'sk-ant-api...'
+}

lib/getChatCompletion/anthropic/config.ts

@@ -1,6 +1,7 @@
-// This isn't very secure, but given only BlueDot Impact employees should have access to the extension and we have a limit on the usage this is fine for now. 
-export const anthropicApiKey = "***REMOVED***";
-// OpenAI model to use, in the format accepted by the OpenAI API: https://platform.openai.com/docs/models/model-endpoint-compatibility
+import { env } from "../../env";
+
+export const anthropicApiKey = env.ANTHROPIC_API_KEY;
+// Model to use: https://docs.anthropic.com/en/docs/models-overview
 export const anthropicModel = 'claude-3-sonnet-20240229';
-// Maximum number of open requests to OpenAI at any one time. Higher = faster, but more likely to hit OpenAI rate limits.
+// Maximum number of open requests at any one time. Higher = faster, but more likely to hit rate limits.
 export const anthropicRequestConcurrency = 30;

lib/getChatCompletion/openai/config.ts

@@ -1,6 +1,8 @@
-// This isn't very secure, but given only BlueDot Impact employees should have access to the extension and we have a limit on the usage this is fine for now. 
-export const openAiApiKey = "***REMOVED***";
-// OpenAI model to use, in the format accepted by the OpenAI API: https://platform.openai.com/docs/models/model-endpoint-compatibility
+import { env } from "../../env";
+
+export const openAiApiKey = env.OPENAI_API_KEY;
+export const openAiOrganisation = "org-2egSVUATOaBoS2Slr2CbYrcZ";
+// Model to use: https://platform.openai.com/docs/models/model-endpoint-compatibility
 export const openAiModel = 'gpt-4-1106-preview';
-// Maximum number of open requests to OpenAI at any one time. Higher = faster, but more likely to hit OpenAI rate limits. From trial and error, at tier 3 with GPT-4 turbo, 30 works well.
+// Maximum number of open requests at any one time. Higher = faster, but more likely to hit rate limits. From trial and error, at tier 3 with GPT-4 turbo, 30 works well.
 export const openAiRequestConcurrency = 30;

lib/getChatCompletion/openai/index.ts

@@ -1,5 +1,5 @@
 import pLimit from 'p-limit';
-import { openAiRequestConcurrency, openAiApiKey, openAiModel } from './config';
+import { openAiRequestConcurrency, openAiApiKey, openAiModel, openAiOrganisation } from './config';
 import { GetChatCompletion } from '..';
 
 const globalRateLimit = pLimit(openAiRequestConcurrency);
@@ -11,7 +11,7 @@ export const getChatCompletion: GetChatCompletion = async (messages) => {
         headers: {
           'Content-Type': 'application/json',
           'Authorization': `Bearer ${openAiApiKey}`,
-          'OpenAI-Organization': 'org-2egSVUATOaBoS2Slr2CbYrcZ',
+          'OpenAI-Organization': openAiOrganisation,
         },
         body: JSON.stringify({
           model: openAiModel,

package.json

@@ -1,5 +1,9 @@
 {
+    "name": "ai-evaluator-extension",
+    "license": "AGPL-3.0-only",
+    "private": true,
     "scripts": {
+        "postinstall": "shx cp -n lib/env.template.ts lib/env.ts",
         "start": "npm run start:applications",
         "start:applications": "block run --remote applications",
         "type-check": "tsc --noEmit --esModuleInterop --jsx react frontend/**.tsx lib/**.ts",
@@ -24,6 +28,7 @@
         "eslint-plugin-react": "^7.31.11",
         "eslint-plugin-react-hooks": "^4.6.0",
         "jest": "^29.3.1",
+        "shx": "^0.3.4",
         "ts-jest": "^29.0.3",
         "typescript": "^4.9.4"
     },