VFP9 and OpenAI 1

[Scott24x7]

Hi All,
I know there was some talk about this a few months back, but can't seem to find it.
I've begun to discover the difference between ChatGPT and OpenAI API.
Anyway, I have an application that has some contact management built into it that I've been using for yonks. It centers around scanning a business card, and populating various data based on that.
I've had a large routine for years that parses the unstructured data on the card (if you've never had to do this, it is surprisingly more complicated than one might expect, especially for such a limited amount of data).
So I got the "bright idea" first to have GPT do the "look at the card" and give me the data. This actually works extremely well. If you take an image (scan) of a card, pass it to ChatGPT it does a perfect job of parsing the data.
So I though, great, let's automate that... BUT, soon discovered that the OpenAI API doesn't accept images for processing.
Bummer.
But then I thought, ok, actually my OCR engine (TOCR) does a fine job of getting the OCR text from the card, so getting the text wasn't so much the issue (ChatGPT does in fact do a better job, especially on cards with colors that are hard to contrast), but my OCR works 98%-ish of the time. So, why not let it do that stage, and then pass THAT to the OpenAPI API call.
Lots of fumbling around with getting an API key generated, and then asking ChatGPT how to integrate this with VFP9 for what I wanted, led me to this point.
Oh, and I should mention, my "prompt" asks OpenAI to parse data passed to it from a business card into specific field titles. (Then I use these to populate the fields with EASE, and I don't need a complex ID routine any longer). Sorry for the long intro explanation to the problem but here is where we are:
ChatGPT created this function for me:

FUNCTION SendToOpenAI
    LPARAMETERS cPrompt, cAPIKey, cModel
    LOCAL oHttp, cURL, cData, cResponse

    * Create the HTTP object
    oHttp = CREATEOBJECT("MSXML2.ServerXMLHTTP.6.0")

    * Determine the correct API endpoint based on the model
    IF cModel $ "gpt-3.5-turbo gpt-4"
        cURL = "https://api.openai.com/v1/chat/completions"
        * Prepare the JSON data for chat-based models
        cData = '{'+;
                '"model": "' + cModel + '",'+;
                '"messages": [{"role": "user", "content": "' + cPrompt + '"}],'+;
                '"max_tokens": 100,'+;
                '"temperature": 0.7'+;
                '}'
    ELSE
        cURL = "https://api.openai.com/v1/completions"
        * Prepare the JSON data for completion models
        cData = '{'+;
                '"model": "' + cModel + '",'+;
                '"prompt": "' + cPrompt + '",'+;
                '"max_tokens": 100,'+;
                '"temperature": 0.7'+;
                '}'
    ENDIF

    * Open a connection and set headers
    oHttp.Open("POST", cURL, .F.)
    oHttp.setRequestHeader("Content-Type", "application/json")
    oHttp.setRequestHeader("Authorization", "Bearer " + cAPIKey)

    * Send the request with data
    oHttp.Send(cData)

    * Check if the response is successful
    IF oHttp.Status = 200
        cResponse = oHttp.ResponseText
        RETURN cResponse
    ELSE
        RETURN "Error: " + oHttp.StatusText + " (" + TRANSFORM(oHttp.Status) + ") - " + oHttp.ResponseText
    ENDIF
ENDFUNC

In fairness, this seems to work (by that I mean, the request is going through, so the API is working).

In my SCAN button, I have this code:

lcInputData =  SendToOpenAI(TTPARSEBC.OCRTEXT, cOpenAPIKey, "gpt-4o")

I wasn't getting what I expected, so I stuck a MESSAGEBOX(lcInputData) call after, and it gives me this:


Now, I note that it's asking about JSON... and what we are doing is passing a fixed prompt (instruction) with variable data. The prompt reads as:

"Use these labels: Company, Honorary, Fname, Mname, Lname, Suffix, Qualifications, Title, Department, URL Mobile, Tel, Fax, AltPhone, email, Building, Addr1, Addr2, City, State, Country, Post, Remains. Where remains is any text that is left over that doesn't fit into any of the labels. Structure the following data, with the labels mentioned and ensure a cr lf is at the end of each. If the item is blank, don't include it."
I then append the OCR text from the card scan to that prompt, so it might end up looking something like this (text from a card I just scanned, but I've scrambled the data so it's not real):
Gaëtan Pennylover, RCDD
JAPAN DISTRICT VICE CHAIR
gpennylover@boxyi.jp
Sagami Bldg 2nd floor, 2-15-6 Ginza
Chuo-ku,
Tokyo 104-0061 Japan
+81.2.5432.8888 / www.boxyi.jp

So this is just sent as a big text string. Now I see something about "JSON", but I've no idea how to pass this as JSON text.
I know this isn't specifically a VFP question, but I suspect others may want to use VFP to do similar work with OpenAI, and perhaps (I've seen some talk about JSON here before), there will be someone who can help me with this?

 

[Chris Miller]

JSON is text data. You do specify oHttp.setRequestHeader("Content-Type", "application/json") and then the only culprit is your request body you provide by oHttp.Send(cData) isn't valid JSON.

You could set your cData with a TEXT..ENDTEXT block instead of a multiline string with several parts, but you should just check whether cData passes a validation to be correct JSON. Details can be usage of quote characters within the prompt text, you would need to escape them with a backslash. I think. Your text could have the wrong codepage, I bet they expect UTF8, not ANSI.

 

[Scott24x7]

I put it in TEXT..ENDTEXT block, and did get a different message, which I think supports your second point. I just had a look, I know I've seen it before but it's been ages. Where can I change the codepage? And is UTF8 even an option in VFP?

 

[Chris Miller]

You have the STRCONV function to convert to UTF8.

Simply as last step before sending cData do

cData = StrConv(cData, 9)

 

[Chris Miller]

Another problem that also was a problem in the old thread ( I found it here: https://www.tek-tips.com/threads/vfp-and-ai-chatgpt.1828484/ )

Don't have line feeds in the cPrompt part, a string in JSON can't have line feeds. You could convert them to \n to state a line feed in the text, it wouldn't matter for the AI to have a long text without any wraps, though, so just remove them.

 

[Scott24x7]

Yeah, I hit that earlier. It seems that UTF-8 issue was an issue, and something had "triggered" me to remove the CR/LFs as well, though I don't recall what.
I have it working now! What an amazing dream.

I had taken months to work out how to extract the data free form (and finding a person's "name" even in a small amount of text is like looking through a sea). This process now takes my OCR'd data, parses it into fields with names that I want, then its a very small CASE statement to just pull them out bit field identifier. I realize now I can shorten them to 1 character id's to save on "tokens" but as it is my round 100ish tests resulted in a charge of 4c. So I'm probably ok. I rarely add more than 50 new cards in a month.

But this outcome is pretty sweet, and now I will tinker a bit with it to fine tune a bit more, and make this even faster. I've also decided to keep the "old method" too, so that if I'm off-line I can still add new cards using my old method, and I don't have to hand enter all the data, it just won't be as accurate. VERY stoked to have this integrated into VFP... I think next it's going to become part of my expense reporting. That it could REALLY boost the time on, but may need a lot more tailoring. I still have the outgoing text in a Text..Endtext block, but not sure if that's even needed now. (That saves another 3 tokens ><).