Show HN: No-code AI-powered API endpoints | Hacker News Reader

Show HN: No-code AI-powered API endpoints

backengine.dev

11 pointssudb5 comments

Backengine is a tool that evolved from an experiment at my startup for which we wanted to pass structured data to an LLM completion API and receive structured data back.

We thought we'd release it as a standalone tool for others who might also find it useful.

Backengine lets you create hosted API endpoints for which the endpoint logic is described in natural language as an LLM prompt. The endpoints accept structured requests and return structured responses.

Why is this useful?

Hosted - no additional infrastructure to maintain.

Prompt engineering workspace - you can edit, test and tweak endpoint prompts without touching any code.

Instant deployment - created endpoints should be immediately available, and changes to your endpoints are deployed in one click.

This is a first and very early release - we'd love for HN to try out Backengine and hear what you think! There's a playground where you can try it out - we're very interested to hear what use-cases people come up with.

Open articleView on HN

Let's get this clear, the endpoint responses are hallucinated by an LLM, right? So this is mostly a tool to experiment with endpoint design in a non-committal way? Or am I missing something fundamental here?

I worked on this with OP.

You instruct the LLM to carry out a task and 'ask' for the result in a structured format.

As always there may be some hallucination which you should be prepared to handle and can minimise with some prompt iteration.

A very simple example here https://blog.backengine.dev/fetching-iso-country-codes-with-...

OK then so the idea is that the LLM writes the response. This is useful for certain types of endpoints (e.g. summarize) but not a generic way to program. Additionally the cost of an LLM eval is currently too high to use this in any high volume use case.

Nevertheless it's quite interesting to think about a future where most API endpoints are "implemented" like this and the "database" they draw from is in the model weigths that underpins a particular API.

Yeah agreed! We reckon there's a fair amount of scope in the kind of endpoints that could be implemented this way - along with summarization, we could also have things like content generation, and parsing of unstructured/semi-structured data.

Great point on cost.

We're kind of assuming cost will come down 10-100x before too long.

Reply on news.ycombinator.com