Added utils for GPT classification using templates and standardizing …

…response. (#252)

services/api/.env

@@ -66,4 +66,7 @@ TWILIO_AUTH_TOKEN=AC21717619d8cf45502cc2f7cd7aee139d
 GOOGLE_CLIENT_ID=
 
 # Sign in with Apple
-APPLE_SERVICE_ID=
+APPLE_SERVICE_ID=
+
+# OpenAI
+OPENAI_API_KEY=

services/api/__mocks__/openai.js

@@ -0,0 +1,21 @@
+let mock;
+
+function OpenAI() {
+  return {
+    chat: {
+      completions: {
+        create() {
+          return mock;
+        },
+      },
+    },
+  };
+}
+
+function setResponse(data) {
+  mock = data;
+}
+
+OpenAI.setResponse = setResponse;
+
+module.exports = OpenAI;

services/api/src/gpt/classify-fruits.md

@@ -0,0 +1,21 @@
+--- SYSTEM ---
+
+You are a helpful assistant.
+
+Here is a list of fruits:
+
+{{fruits}}
+
+--- USER ---
+
+The following text describes someone eating a meal. Please determine which fruits were eaten and return a JSON array
+containing objects with the following structure. Only output JSON, do not include any explanations.
+
+- "name" - The name of the fruit.
+- "color" - The typical color of the fruit.
+- "calories" - A rough estimate of the number of calories per serving. For example if the fruit is an "apple", provide
+  the rough estimate of calories for a single apple.
+
+Text:
+
+{{text}}

services/api/src/utils/__fixtures__/gpt-response-array-unformatted.json

@@ -0,0 +1,23 @@
+{
+  "id": "chatcmpl-9dy8si0kRlF27OZiDtA4Y38u4lfO1",
+  "object": "chat.completion",
+  "created": 1719313006,
+  "model": "gpt-4o-2024-05-13",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "[{\n\"name\": \"banana\",\n  \"color\": \"yellow\",\n  \"calories\": 105\n}]"
+      },
+      "logprobs": null,
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 134,
+    "completion_tokens": 79,
+    "total_tokens": 213
+  },
+  "system_fingerprint": "fp_3e7d703517"
+}

services/api/src/utils/__fixtures__/gpt-response-formatted.json

@@ -0,0 +1,23 @@
+{
+  "id": "chatcmpl-9dy8si0kRlF27OZiDtA4Y38u4lfO1",
+  "object": "chat.completion",
+  "created": 1719313006,
+  "model": "gpt-4o-2024-05-13",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "```json\n{\n  \"name\": \"banana\",\n  \"color\": \"yellow\",\n  \"calories\": 105\n}\n```"
+      },
+      "logprobs": null,
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 134,
+    "completion_tokens": 79,
+    "total_tokens": 213
+  },
+  "system_fingerprint": "fp_3e7d703517"
+}

services/api/src/utils/__fixtures__/gpt-response-long.json

@@ -0,0 +1,23 @@
+{
+  "id": "chatcmpl-9dy8si0kRlF27OZiDtA4Y38u4lfO1",
+  "object": "chat.completion",
+  "created": 1719313006,
+  "model": "gpt-4o-2024-05-13",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "To determine the fruit eaten, we analyze the text provided. Here, the person ate a banana. We'll now create a JSON object including the name, typical color, and a rough estimate of calories per serving for a banana.\n\nHere is the JSON object:\n\n```json\n{\n  \"name\": \"banana\",\n  \"color\": \"yellow\",\n  \"calories\": 105\n}\n```"
+      },
+      "logprobs": null,
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 134,
+    "completion_tokens": 79,
+    "total_tokens": 213
+  },
+  "system_fingerprint": "fp_3e7d703517"
+}

services/api/src/utils/__fixtures__/gpt-response-unformatted.json

@@ -0,0 +1,23 @@
+{
+  "id": "chatcmpl-9dy8si0kRlF27OZiDtA4Y38u4lfO1",
+  "object": "chat.completion",
+  "created": 1719313006,
+  "model": "gpt-4o-2024-05-13",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": "{\n  \"name\": \"banana\",\n  \"color\": \"yellow\",\n  \"calories\": 105\n}"
+      },
+      "logprobs": null,
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 134,
+    "completion_tokens": 79,
+    "total_tokens": 213
+  },
+  "system_fingerprint": "fp_3e7d703517"
+}

services/api/src/utils/__tests__/gpt.js

@@ -0,0 +1,139 @@
+const { prompt } = require('../gpt');
+const { setResponse } = require('openai');
+
+const responseLong = require('../__fixtures__/gpt-response-long.json');
+const responseFormatted = require('../__fixtures__/gpt-response-formatted.json');
+const responseUnformatted = require('../__fixtures__/gpt-response-unformatted.json');
+const responseArrayUnformatted = require('../__fixtures__/gpt-response-array-unformatted.json');
+
+describe('prompt', () => {
+  it('should succeed for a long response', async () => {
+    setResponse(responseLong);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+    });
+    expect(result).toEqual({
+      name: 'banana',
+      color: 'yellow',
+      calories: 105,
+    });
+  });
+
+  it('should succeed for a formatted response', async () => {
+    setResponse(responseFormatted);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+    });
+    expect(result).toEqual({
+      name: 'banana',
+      color: 'yellow',
+      calories: 105,
+    });
+  });
+
+  it('should succeed for an unformatted response', async () => {
+    setResponse(responseUnformatted);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+    });
+    expect(result).toEqual({
+      name: 'banana',
+      color: 'yellow',
+      calories: 105,
+    });
+  });
+
+  it('should succeed for an array response', async () => {
+    setResponse(responseArrayUnformatted);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+    });
+    expect(result).toEqual([
+      {
+        name: 'banana',
+        color: 'yellow',
+        calories: 105,
+      },
+    ]);
+  });
+
+  it('should be able to return all messages', async () => {
+    setResponse(responseArrayUnformatted);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+      output: 'messages',
+    });
+    expect(result).toEqual([
+      {
+        role: 'system',
+        content: 'You are a helpful assistant.\n\nHere is a list of fruits:',
+      },
+      {
+        role: 'user',
+        content:
+          'The following text describes someone eating a meal. Please determine which fruits were eaten and return a JSON array\n' +
+          'containing objects with the following structure. Only output JSON, do not include any explanations.\n' +
+          '\n' +
+          '- "name" - The name of the fruit.\n' +
+          '- "color" - The typical color of the fruit.\n' +
+          '- "calories" - A rough estimate of the number of calories per serving. For example if the fruit is an "apple", provide\n' +
+          '  the rough estimate of calories for a single apple.\n' +
+          '\n' +
+          'Text:\n' +
+          '\n' +
+          'I had a burger and some french fries for dinner. For dessert I had a banana.',
+      },
+      {
+        role: 'assistant',
+        content: '[{\n"name": "banana",\n  "color": "yellow",\n  "calories": 105\n}]',
+      },
+    ]);
+  });
+
+  it('should be able to output just the text', async () => {
+    setResponse(responseArrayUnformatted);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+      output: 'text',
+    });
+    expect(result).toBe('[{\n"name": "banana",\n  "color": "yellow",\n  "calories": 105\n}]');
+  });
+
+  it('should be able to output the raw response', async () => {
+    setResponse(responseArrayUnformatted);
+    const result = await prompt({
+      file: 'classify-fruits',
+      text: 'I had a burger and some french fries for dinner. For dessert I had a banana.',
+      output: 'raw',
+    });
+    expect(result).toEqual({
+      id: 'chatcmpl-9dy8si0kRlF27OZiDtA4Y38u4lfO1',
+      object: 'chat.completion',
+      created: 1719313006,
+      model: 'gpt-4o-2024-05-13',
+      choices: [
+        {
+          index: 0,
+          message: {
+            role: 'assistant',
+            content: '[{\n"name": "banana",\n  "color": "yellow",\n  "calories": 105\n}]',
+          },
+          logprobs: null,
+          finish_reason: 'stop',
+        },
+      ],
+      usage: {
+        prompt_tokens: 134,
+        completion_tokens: 79,
+        total_tokens: 213,
+      },
+      system_fingerprint: 'fp_3e7d703517',
+    });
+  });
+});

services/api/src/utils/gpt.js

@@ -0,0 +1,94 @@
+const fs = require('fs/promises');
+const path = require('path');
+const OpenAI = require('openai');
+const Mustache = require('mustache');
+const config = require('@bedrockio/config');
+const { memoize } = require('lodash');
+
+const openai = new OpenAI({
+  apiKey: config.get('OPENAI_API_KEY'),
+});
+
+const TEMPLATE_DIR = path.join(__dirname, '../gpt');
+const MESSAGES_REG = /(?:^|\n)-{3,}\s*(\w+)\s*-{3,}(.*?)(?=\n-{3,}|$)/gs;
+const JSON_REG = /([{[].+[}\]])/s;
+
+const MODEL = 'gpt-4o';
+
+async function prompt(options) {
+  const messages = await getMessages(options);
+  return await runCompletion(messages, options);
+}
+
+async function getMessages(options) {
+  const { file, output, ...rest } = options;
+  const template = await loadTemplate(file);
+
+  const raw = Mustache.render(template, transformParams(rest));
+
+  const messages = [];
+  for (let match of raw.matchAll(MESSAGES_REG)) {
+    const [, role, content] = match;
+    messages.push({
+      role: role.toLowerCase(),
+      content: content.trim(),
+    });
+  }
+
+  return messages;
+}
+
+function transformParams(params) {
+  const result = {};
+  for (let [key, value] of Object.entries(params)) {
+    if (Array.isArray(value)) {
+      value = value
+        .map((el) => {
+          return `- ${el}`;
+        })
+        .join('\n');
+    } else if (typeof value === 'object') {
+      value = JSON.stringify(value, null, 2);
+    }
+    result[key] = value;
+  }
+  return result;
+}
+
+async function runCompletion(messages, options) {
+  const { output = 'json' } = options;
+
+  const response = await openai.chat.completions.create({
+    model: MODEL,
+    messages,
+  });
+
+  let content = response.choices[0].message.content;
+
+  if (output === 'raw') {
+    return response;
+  } else if (output === 'text') {
+    return content;
+  } else if (output === 'messages') {
+    const { message } = response.choices[0];
+    return [...messages, message];
+  } else if (output === 'json') {
+    try {
+      const match = content.match(JSON_REG);
+      return JSON.parse(match[1]);
+    } catch (error) {
+      throw new Error('Unable to derive JSON object in response.');
+    }
+  }
+}
+
+const loadTemplate = memoize(async (file) => {
+  if (!file.endsWith('.md')) {
+    file += '.md';
+  }
+  return await fs.readFile(path.join(TEMPLATE_DIR, file), 'utf8');
+});
+
+module.exports = {
+  prompt,
+};