Add CLI option to output computation as structured format

[aurelleb]

Hi there, I'm looking to make numbat available as a calculator backend for https://github.com/vicinaehq/vicinae and I'm hitting a roadblock trying to integrate it.

Since I'm dealing with a C++ codebase I can't use the library directly, so I thought I could just invoke numbat through its CLI every time a computation is required.

I also considered using FFI but honestly that sounds like a lot of work only to integrate a calculator backend.

The problem is, the output is not meant to be programatically parsed, making this a very hard to implement solution. Errors in particular, cannot be easily parsed at all.

So my suggestion would be to implement a flag that would output the result as a structured format, typically JSON, something like this:

numbat -e '10km to m' --color=never --json

with as an output:

{
  "value": "10000",
  "unit": "m"
}

or

{
  "error": "Incompatible units"
}

The output could be a bit more detailed, I don't know.

I think this would make integrating numbat with launchers or even scripts significantly easier.

If you think this has its place upstream then I'm willing to make a PR for this.

---

On a side note, it would also be great if we could override specific config options when using the CLI with the -e flag.

In order to make CLI usage a viable option for these kind of use cases we would probably need to add a new exchange rate loading mode that caches it on disk instead of fetching them anew every invocation. But I diverge, that should probably be its own issue.

[leiserfg]

I think also exposing the "expanded expression" would be nice, as it helps the user to be sure numbat is understanding its intention. For the same query as op, something like:

{
   "expanded" : "10 kilometre ➞ metre",
   "value": "1000",
   "unit": "m",
  "dimension": "Length"
}

[Goju-Ryu]

I think this would be a good addition to numbat. Feel free to open a PR for it.

These two issues seem partially related (although parsing focused): #470 and #262
The discussions there may provide some inspiration.

[leiserfg]

@aurelien-brabant for vicinae wouldn't it work as well if instead of running the cli every time, a session of numbat was kept running and it would be possible to talk to it with a JSON-RPC over stdio? Given that numbat has a repl, that does not seem very difficult to implement (in this side). That should lead to quicker responses and maybe even remove the need for caching the exchange rates.

[aurelleb]

that's another option but having to implement a json-rpc interface is more complex, not sure numbat maintainers want that?

[sharkdp]

You could probably start experimenting with this by sending something like the following to the numbat CLI, instead of just the user-expression (10km to m):

let user_expr = # actual user input here

let out_value = value_of(user_expr)
let out_unit = unit_name(user_expr)

print("{{\"value\": \"{out_value}\", \"unit\": \"{out_unit}\"}}")

(try here)

I think also exposing the "expanded expression" would be nice, as it helps the user to be sure numbat is understanding its intention. For the same query as op, something like:

{ "expanded" : "10 kilometre ➞ metre", "value": "1000", "unit": "m", "dimension": "Length" }

I agree, this would certainly be more valuable. I'm not opposed to adding a JSON output option, but it would certainly need some design work first. For example: What would this mode do for more complex (non-single-expression) Numbat programs? What happens if there are print statements in the Numbat program?

[leiserfg]

@sharkdp I think it should not print and return only the json result for the last expression.

[Goju-Ryu]

@sharkdp I think it should not print and return only the json result for the last expression.

I think returning a list of results may be a good idea. In cases where you give a script with multiple outputs, it could be valuable to be able to get all of the results. That way it is also possible to get no outputs if a script for example consists only of print statements (assuming we went with their output not being included).

[leiserfg]

Guys any output about doing JSON-RPC over stdio? I'm thinking of one with two methods: pseudo-code

eval("string") -> vec<Output> reloadCurrencies() -> bool

where Output has a kind to identify it and can be either an error, an expression result and a procedure call (like for print).

[leiserfg]

I just noticed @irevoire is working in a lsp, that might do it 🤞

[irevoire]

Huuum I don't think so, the lsp is not interpreting your code and doesn't load the currency at all 😬

[leiserfg]

But it could, it could have a workspace/executeCommand to load currency and a codeLens to get the evaluation of a chunk, for instance.