Doesn't the command-line utility `jq` already define a protocol for this? How do the syntaxes compare?
(LLMs are already very adept at using `jq` so I would think it was preferable to be able to prompt a system that implements querying inside of source code as "this command uses the same format as `jq`")
Oh wow, it got undeleted. Some editor insisted on deleting it because it was a "personal project" (Stephen Dolan's) even though it has a huge user base. I guess now that it has a proper "org" in GitHub it's different. What nonsense.
I read the man page of `jq` and learned how to use it. It's quite well-written and contains a good introduction.
I've observed that too many users of jq aren't willing to take a few minutes to understand how stream programming works. That investment pays off in spades.
Also, LLMs are good at spitting out filters, but you can learn what they do by going and then looking up what it’s doing in the docs. They often apply things in far more interesting and complex ways than the docs at jqlang.org do, which are often far too “foo bar baz” tier to truly understand explain the power of things.
Maybe the author would be in a better place to do that, having the expertise already. Also, as a user I'm quite happy with jq already, so why expend the effort?
I have a similar use case in the app I'm working on. Initially I went with JSONata, which worked, but resulted in queries that indeed felt more like incantations and were difficult even for me to understand (let alone my users).
I then switched to JavaScript / TypeScript, which I found much better overall: it's understandable to basically every developer, and LLMs are very good at it. So now in my app I have a button wherever a TypeScript snippet is required that asks the LLM for its implementation, and even "weak" models one-shot it correctly 99% of the times.
It's definitely more difficult to set up, though, as it requires a sandbox where you can run the code without fears. In my app I use QuickJS, which works very well for my use case, but might not be performant enough in other contexts.
You just have to wrap your mind around jq. It's a) functional, b) has pervasive generators and backtracking. So when you write `.a[].b`, which is a lot like `(.a | .[] | .b)` what you get is three generators strung together in an `and_then` fashion: `.a`, then `.[]`, and then `.b`. And here `.a` generates exactly one value, as does `.b`, but `.[]` generates as many values as are in the value produced by `.a`. And obviously `.b` won't run at all if `.a` has no values, and `.b` will run for _each_ value of `.a[]`. Once you begin to see the generators and the backtracking then everything begins to make sense.
Most alternatives being talked about are working on query strings (like `$.phoneNumbers[:1].type`) which is fine but can not be easily modeled / modified by code.
Things like https://jsonlogic.com/ works better if you wish to expose a rest api with a defined query schema or something like that. Instead of accepting a query `string`. This seems better as in you have a string format and a concrete JSON format. Also APIs to convert between them.
Also, if you are building a filter interface, having a structured representation helps:
Cool idea! Although without looking closer I can't tell if "meme" is in reference to the technical or the colloquial meaning of meme.
Admittedly I don't know that much about LLM optimization/configuration, so apologies if I'm asking dumb questions. Isn't the value of needing to copy/paste that prompt in front of your queries a huge bog on net token efficiency? Like wouldn't you need to do some hundred/thousand query translations just to break even? Maybe I don't understand what you've built.
Thank you. That script prompt is just for development and exploration. A production model needs to be trained/fine-tuned on Memelang first. We're working on this now. The math says we can deliver a model 1/2 the size of an equivalent model for SQL.
This is all too cute. Why not just use JavaScript syntax? You can limit it to the exact amount of functionality you want for whatever reason it is you want to limit it.
If you prefer JSONPath as a query language, oj from https://github.com/ohler55/ojg provides that functionality. It can also be installed with brew. (disclaimer, I'm the author of OjG)
Helpful when querying JSON API responses that are parsed and persisted for normal, relational uses. Sometimes you want to query data that you weren’t initially parsing or that matches a fix to reprocess.
speaking of classic databases: can anyone explain to me, a dummy, why any syntax like this or even GraphQL is preferable to "select a.name, a.age from friends a where a.city = 'New York' order by a.age asc"?
There is a standard in RFC 9535 (JSONPath)[1]. But as far as I can tell, it isn't very widely used, and it has more limited functionality than some of the alternatives.
the issue with JSONPath is that it took 17 years for it to become a properly fleshed-out standard. The original idea came from a 2007 blog post [0], which was then extended and implemented subtly differently dozens of times, with the result that almost every JSON Path implementation out there is incompatible with the others.
Plus, I feel like most, if not all, higher level languages already come with everything you need to do that easily. Well except for go that requires you to create your own filter function.
The AWS CLI supports JMESPath (https://jmespath.org) for the `--query` flag. I don't think I've run into anything else that uses it. Pretty similar to JSONPath IIRC.
jq is good but its syntax is strangely unmemorizable. Have used it for a decade and always need to look at the manual or at examples to refresh my knowledge.
I use `jsonata` currently at work. I think it's excellent. There's even a limited-functionality rustlib (https://github.com/Stedi/jsonata-rs). What I particularly like about `jsonata` is its support for variables, they're super useful in a pinch when a pure expression becomes ugly or unwieldy or redundant. It also lets you "bring your own functions", which lets you do things like:
```
$sum($myArrayExtractor($.context))
```
where `$myArrayExtractor` is your custom code.
---
Re: "how did it go"
We had a situation where we needed to generate EDI from json objects, which routinely required us to make small tweaks to data, combine data, loop over data, etc. JSONata provided a backend framework for data transformations that reduced the scope and complexity of the project drastically.
I think JSONata is an excellent fit for situations where companies need to do data transforms, for example when it's for the sake of integrations from 3rd-party sources; all the data is there, it just needs to be mapped. Instead of having potentially buggy code as integration, you can have a pseudo-declarative jsonata spec that describes the transform for each integration source, and then just keep a single unified "JSONata runner" as the integration handler.
It's nice because we can just put the JSONata expression into a db field, and so you can have arbitrary data transforms for different customers for different data structures coming or going, and they can be set up just by editing the expression via the site, without having to worry about sandboxing it (other than resource exhaustion for recursive loops). It really sped up the iteration process for configuring transforms.
I hate jq as much as the next guy but it’s ubiquitous and great for this sort of thing. If you want a single path style query language I’d highly recommend JsonPath. It’s so much nicer than jq for “I need every student’s gpa”.
Doesn't the command-line utility `jq` already define a protocol for this? How do the syntaxes compare?
(LLMs are already very adept at using `jq` so I would think it was preferable to be able to prompt a system that implements querying inside of source code as "this command uses the same format as `jq`")
Mongo also has a good query language and a mongo DB can be seen as an array of documents
For convenience: https://en.wikipedia.org/wiki/Jq_(programming_language)
Oh wow, it got undeleted. Some editor insisted on deleting it because it was a "personal project" (Stephen Dolan's) even though it has a huge user base. I guess now that it has a proper "org" in GitHub it's different. What nonsense.
Maybe it helped that they called it a "programming language"? It helps make it sound super serious.
Wikipedia is such a disappointment
DuckDB can read JSON - you can query JSON with normal SQL.[1] I prefer to Malloy Data language for querying as it is 10x simpler than SQL.[2]
[1] - https://duckdb.org/docs/stable/data/json/overview [2] - https://www.malloydata.dev/
So can postgres, I tend to just use PG, since I have instances running basically everywhere, even locally, but duckdb works well too.
I read the man page of `jq` and learned how to use it. It's quite well-written and contains a good introduction.
I've observed that too many users of jq aren't willing to take a few minutes to understand how stream programming works. That investment pays off in spades.
Also, LLMs are good at spitting out filters, but you can learn what they do by going and then looking up what it’s doing in the docs. They often apply things in far more interesting and complex ways than the docs at jqlang.org do, which are often far too “foo bar baz” tier to truly understand explain the power of things.
I'd like to know how it compares to https://jsonata.org
JSONata looks to be more general purpose with its support for variables/statements, and custom functions. I'd probably still stick with JSONata
Can't you just visit both pages, build an understanding and compare them?
Maybe the author would be in a better place to do that, having the expertise already. Also, as a user I'm quite happy with jq already, so why expend the effort?
I have a similar use case in the app I'm working on. Initially I went with JSONata, which worked, but resulted in queries that indeed felt more like incantations and were difficult even for me to understand (let alone my users).
I then switched to JavaScript / TypeScript, which I found much better overall: it's understandable to basically every developer, and LLMs are very good at it. So now in my app I have a button wherever a TypeScript snippet is required that asks the LLM for its implementation, and even "weak" models one-shot it correctly 99% of the times.
It's definitely more difficult to set up, though, as it requires a sandbox where you can run the code without fears. In my app I use QuickJS, which works very well for my use case, but might not be performant enough in other contexts.
You just have to wrap your mind around jq. It's a) functional, b) has pervasive generators and backtracking. So when you write `.a[].b`, which is a lot like `(.a | .[] | .b)` what you get is three generators strung together in an `and_then` fashion: `.a`, then `.[]`, and then `.b`. And here `.a` generates exactly one value, as does `.b`, but `.[]` generates as many values as are in the value produced by `.a`. And obviously `.b` won't run at all if `.a` has no values, and `.b` will run for _each_ value of `.a[]`. Once you begin to see the generators and the backtracking then everything begins to make sense.
In the k8s world there's a random collection of json path, json query, some random expression language.
Just use jq. None of the other ones are as flexible or widespread and you just end up with frustrated users.
This. Jq is the defacto standard and anytime I come across something else I am annoyed.
Which isn't to say jq is the best or even good but its battle-tested and just about every conceivable query problem has been thrown at it by now.
Most alternatives being talked about are working on query strings (like `$.phoneNumbers[:1].type`) which is fine but can not be easily modeled / modified by code.
Things like https://jsonlogic.com/ works better if you wish to expose a rest api with a defined query schema or something like that. Instead of accepting a query `string`. This seems better as in you have a string format and a concrete JSON format. Also APIs to convert between them.
Also, if you are building a filter interface, having a structured representation helps:
https://react-querybuilder.js.org/demo?outputMode=export&exp...
JSON logic is nice, but for example, the Python bindings were last updated 8 years ago
I can't help myself and surely someone else has already done the same. But the query
does exactly the same without any plugin.am I missing something?
The verbosity.
To your point abstractions often multiply and then hide the complexity, and create a facade of simplicity.
I've been working on an ultra-token-efficient LLM-friendly query language. https://memelang.net/09/
Cool idea! Although without looking closer I can't tell if "meme" is in reference to the technical or the colloquial meaning of meme.
Admittedly I don't know that much about LLM optimization/configuration, so apologies if I'm asking dumb questions. Isn't the value of needing to copy/paste that prompt in front of your queries a huge bog on net token efficiency? Like wouldn't you need to do some hundred/thousand query translations just to break even? Maybe I don't understand what you've built.
Cool idea either way!
Thank you. That script prompt is just for development and exploration. A production model needs to be trained/fine-tuned on Memelang first. We're working on this now. The math says we can deliver a model 1/2 the size of an equivalent model for SQL.
.friends | filter(.city == "New York") | sort(.age) | pick(.name, .age)
mapValues(mapKeys(substring(get(), 0, 10)))
This is all too cute. Why not just use JavaScript syntax? You can limit it to the exact amount of functionality you want for whatever reason it is you want to limit it.
If you prefer JSONPath as a query language, oj from https://github.com/ohler55/ojg provides that functionality. It can also be installed with brew. (disclaimer, I'm the author of OjG)
JSONPath is also supported by Postgres!
Helpful when querying JSON API responses that are parsed and persisted for normal, relational uses. Sometimes you want to query data that you weren’t initially parsing or that matches a fix to reprocess.
speaking of classic databases: can anyone explain to me, a dummy, why any syntax like this or even GraphQL is preferable to "select a.name, a.age from friends a where a.city = 'New York' order by a.age asc"?
jq is amazing (not only for querying json), I recommend going though the docs, it's fairly small.
I implemented one day of advent of code in jq to learn it: https://github.com/ivanjermakov/adventofcode/blob/master/aoc...
There are a ridiculous number of JSON query/path languages. Wish all the authors got together and harmonized on a standard.
There is a standard in RFC 9535 (JSONPath)[1]. But as far as I can tell, it isn't very widely used, and it has more limited functionality than some of the alternatives.
[1]: https://datatracker.ietf.org/doc/html/rfc9535
the issue with JSONPath is that it took 17 years for it to become a properly fleshed-out standard. The original idea came from a 2007 blog post [0], which was then extended and implemented subtly differently dozens of times, with the result that almost every JSON Path implementation out there is incompatible with the others.
[0] https://goessner.net/articles/JsonPath/
Don't forget the also standardized way of referring to a single value in JSON, "JSON Pointer": datatracker.ietf.org/doc/html/rfc6901
Postgresql supports jsonpath, right?
SQLite might too, though I'm struggling to find anything explicit about the syntax: https://sqlite.org/json1.html#jptr
it might just be a very limited subset?
Looks like "JSON Pointer": datatracker.ietf.org/doc/html/rfc6901
Plus, I feel like most, if not all, higher level languages already come with everything you need to do that easily. Well except for go that requires you to create your own filter function.
The AWS CLI supports JMESPath (https://jmespath.org) for the `--query` flag. I don't think I've run into anything else that uses it. Pretty similar to JSONPath IIRC.
azure tools also support JMESPath
The standard is called jq, any new standard is just going to be a committee circle jerk that doesn't move the ball forward in any meaningful way.
jq is good but its syntax is strangely unmemorizable. Have used it for a decade and always need to look at the manual or at examples to refresh my knowledge.
Xkcd.gif
https://jsonpath.com/ or https://jsonata.org/
Would you mind sharing a bit more? Have you used them? How did that go?
I use `jsonata` currently at work. I think it's excellent. There's even a limited-functionality rustlib (https://github.com/Stedi/jsonata-rs). What I particularly like about `jsonata` is its support for variables, they're super useful in a pinch when a pure expression becomes ugly or unwieldy or redundant. It also lets you "bring your own functions", which lets you do things like:
``` $sum($myArrayExtractor($.context)) ```
where `$myArrayExtractor` is your custom code.
---
Re: "how did it go"
We had a situation where we needed to generate EDI from json objects, which routinely required us to make small tweaks to data, combine data, loop over data, etc. JSONata provided a backend framework for data transformations that reduced the scope and complexity of the project drastically.
I think JSONata is an excellent fit for situations where companies need to do data transforms, for example when it's for the sake of integrations from 3rd-party sources; all the data is there, it just needs to be mapped. Instead of having potentially buggy code as integration, you can have a pseudo-declarative jsonata spec that describes the transform for each integration source, and then just keep a single unified "JSONata runner" as the integration handler.
We've had a great experience with JSONata too.
It's nice because we can just put the JSONata expression into a db field, and so you can have arbitrary data transforms for different customers for different data structures coming or going, and they can be set up just by editing the expression via the site, without having to worry about sandboxing it (other than resource exhaustion for recursive loops). It really sped up the iteration process for configuring transforms.
What do your users know? If they’re quite familiar with SQL for querying, I would look at duckdb.
Maybe JS directly?
Nice work with a jq-esque feel. Website is cut on mobile devices though
Nice. I work on something similar but for .net.
They have an implementation for .net https://jsonquerylang.org/implementations/#net
I hate jq as much as the next guy but it’s ubiquitous and great for this sort of thing. If you want a single path style query language I’d highly recommend JsonPath. It’s so much nicer than jq for “I need every student’s gpa”.
not to be confused with jq for querying json?