invoke_model_with_response_stream generates lexical error: invalid char in json text, but points to blank

[tonisama]

The invoke_model_with_response_stream function generates an error:

Error in parse_con(txt, bigint_as_char) :
lexical error: invalid char in json text.
(right here) ------^

The error occurs at the response request. The (right here) usually points to the infringing character, but points to nothing here. I replicated this same script in Python using the boto3 library and it works perfectly. In R, the invoke_model function without the streaming also works perfectly.

To ensure I wasn't making an error in the prompt encoding, I copied the hex representation from Python into R and converted it back to string which still gave me the same error.

Unfortunately, due to the error, R doesn't record the response, so I'm unable to view any server-side error messages.

[DyfanJones]

HI @tonisama can you provide a code example please. To help me reproduce the error.

[tonisama]

HI @tonisama can you provide a code example please. To help me reproduce the error.

Sure, but you'll need bedrock credentials to use it.
bedrock-test-stream-v2.txt

[DyfanJones]

This error is coming from

paws/paws.common/R/jsonutil.R

Lines 4 to 17 in 45ece0e

	decode_json <- function(raw) {
	if (length(raw) == 0) {
	return(list())
	}
	con <- rawConnection(raw)
	on.exit(close(con))
	obj <- jsonlite::fromJSON(
	con,
	simplifyVector = FALSE,
	simplifyDataFrame = FALSE,
	simplifyMatrix = FALSE
	)
	return(obj)
	}

I am not sure what is going on regarding the parsing of the body. However I believe this method is not fully supported as of yet. Paws will need to develop a streaming method to parse the data. I will have to see how botocore does it and go from there. In the meantime is invoke_model method suitable?

[DyfanJones]

As paws is currently built off httr. We will have to investigate this method: https://httr.r-lib.org/reference/write_stream.html or potentially go to httr2 or the underlying curl package.

[tonisama]

This error is coming from

paws/paws.common/R/jsonutil.R

Lines 4 to 17 in 45ece0e

	decode_json <- function(raw) {
	if (length(raw) == 0) {
	return(list())
	}
	con <- rawConnection(raw)
	on.exit(close(con))
	obj <- jsonlite::fromJSON(
	con,
	simplifyVector = FALSE,
	simplifyDataFrame = FALSE,
	simplifyMatrix = FALSE
	)
	return(obj)
	}

I am not sure what is going on regarding the parsing of the body. However I believe this method is not fully supported as of yet. Paws will need to develop a streaming method to parse the data. I will have to see how botocore does it and go from there. In the meantime is invoke_model method suitable?

Thanks! invoke_model works just fine.