Repository files navigation

Basic caching proxy for OpenAI API, deployable as aCloudflare Worker.

This can help you reduce your OpenAI costs (and get faster results) by returning cached responses for repeated requests.

The proxy server supports specifying cache TTL on a per-request basis, so you could configure this based on your needs. For example, thetext-davinci-003 model is 10x the cost oftext-curie-001 so you could choose to cache results for longer for davinci.

Client compatibility:

• openai/openai-node: full compatibility, takes just a few lines of config to use
• openai/openai-python: partial compatibility, supports caching but no TTL options so you'll need a cache eviction policy
• alexrudall/ruby-openai: partial compatibility, supports caching but no TTL options so you'll need a cache eviction policy

It only cachesPOST requests that have a JSON request body, as these tend to be the slowest and are the only ones that cost money (for now).

Clone the repo and install dependencies.

You will need to sign up for two services (which both have free tiers):

• Cloudflare: Where our worker will be hosted.
• Upstash: We use Upstash's redis-over-HTTP service for storing cached OpenAI responses.

Finally, set up your redis secrets based on instructions inwrangler.toml.

Depending on your usage, you may try replacing Redis withCloudflare KV instead which is eventually consistent but will likely provide better read latency. Checkwrangler.toml for setup instructions.

Start the proxy server athttp://localhost:8787 with:

yarn start

Then, in your separate project where you have youropenai/openai-node configuration, pass in the newbasePath so that it sends requests through your proxy rather than directly to OpenAI:

const { Configuration, OpenAIApi } = require("openai");const configuration = new Configuration({  apiKey: process.env.OPENAI_API_KEY,+ // Point this to your local instance or Cloudflare deployment:+ basePath: 'http://localhost:8787/proxy',});const openai = new OpenAIApi(configuration);

You can then try a few sample requests. The first will be proxied to OpenAI since a cached response isn't yet saved for it, but the second repeated/duplicate request will return the cached result instead.

constoptions={model:'text-ada-001',prompt:'write a poem about computers'};// This first request will be proxied as-is to OpenAI API, since a cached// response does not yet exist for it:constcompletion=awaitopenai.createCompletion(options);console.log('completion:',completion);// This second request uses the same options, so it returns nearly instantly from// local cache and does not make a request to OpenAI:constcompletionCached=awaitopenai.createCompletion(options);console.log('completionCached:',completionCached);

If you don't want to indefinitely cache results, or you don't have an eviction policy set up on your redis instance, you can specify a TTL in seconds using theX-Proxy-TTL header.

const configuration = new Configuration({  ...+ baseOptions: {+   // In this example, we specify a cache TTL of 24 hours before it expires:+   headers: { 'X-Proxy-TTL': 60 * 60 * 24 }+ }});

If you need to force refresh the cache, you can use the headerX-Proxy-Refresh. This will fetch a new response from OpenAI and cache this new response.

const configuration = new Configuration({  ...+ baseOptions: {+   headers: { 'X-Proxy-Refresh': 'true' }+ }});

See/examples/ directory for a full example of how to call this proxy with your openai client.

This includes both Node.js, Python and Ruby client usage examples.