|
| 1 | +# Input and Output Filters (Advanced) |
| 2 | + |
| 3 | +GPTScript supports input and output filters, which are tools that can modify the input to a tool or the output from a tool. |
| 4 | +These are best explained with examples. |
| 5 | + |
| 6 | +## Input Filter Example |
| 7 | + |
| 8 | +In this example, the entrypoint tool uses an input filter to modify the `message` parameter, before calling the subtool. |
| 9 | +Then, the subtool uses another input filter to modify the message, then writes it to a file. |
| 10 | + |
| 11 | +``` |
| 12 | +# File name: script.gpt |
| 13 | +Param: message: the message from the user |
| 14 | +Tools: subtool |
| 15 | +Input Filter: appleToOrange |
| 16 | +
|
| 17 | +Take the message and give it to the subtool. Then say "Done". |
| 18 | +
|
| 19 | +--- |
| 20 | +Name: subtool |
| 21 | +Param: message: the message from the user |
| 22 | +Input Filter: orangeToBanana |
| 23 | +
|
| 24 | +#!python3 |
| 25 | +
|
| 26 | +import os |
| 27 | +
|
| 28 | +message = os.getenv("message", "") |
| 29 | +with open("gptscript_output.txt", "w") as f: |
| 30 | + f.write(message) |
| 31 | +
|
| 32 | +--- |
| 33 | +Name: appleToOrange |
| 34 | +
|
| 35 | +#!python3 |
| 36 | +
|
| 37 | +import os |
| 38 | +
|
| 39 | +def output(input: str): |
| 40 | + return input.replace("apple", "orange") |
| 41 | +
|
| 42 | +print(output(os.getenv("INPUT", ""))) |
| 43 | +
|
| 44 | +--- |
| 45 | +Name: orangeToBanana |
| 46 | +
|
| 47 | +#!python3 |
| 48 | +
|
| 49 | +import os |
| 50 | +
|
| 51 | +def output(input: str): |
| 52 | + return input.replace("orange", "banana") |
| 53 | + |
| 54 | +print(output(os.getenv("INPUT", ""))) |
| 55 | +``` |
| 56 | + |
| 57 | +Try running this tool with the following command: |
| 58 | + |
| 59 | +```bash |
| 60 | +gptscript script.gpt '{"message":"apple is great"}' |
| 61 | + |
| 62 | +# Then view the output: |
| 63 | +cat gptscript_output.txt |
| 64 | +``` |
| 65 | + |
| 66 | +The output should say "banana is great". |
| 67 | +This matches what we expect, because the input filter `appleToOrange` changes "apple" to "orange", |
| 68 | +and the input filter `orangeToBanana` changes "orange" to "banana". |
| 69 | +If we run the tool again with a different message, like "hello world", the final message will be unmodified, |
| 70 | +since it did not include the words "apple" or "orange". |
| 71 | + |
| 72 | +The input filter tools both read the input from the environment variable `INPUT`. |
| 73 | +They write their modified input to stdout. |
| 74 | +This variable is set by GPTScript before running the input filter tool. |
| 75 | + |
| 76 | +### Input Filter Real-World Example |
| 77 | + |
| 78 | +For a real-world example of an input filter tool, check out the [gptscript-ai/context/at-syntax](https://github.com/gptscript-ai/context/tree/main/at-syntax) tool. |
| 79 | + |
| 80 | +## Output Filter Example |
| 81 | + |
| 82 | +In this example, the tool is asked to write a poem about apples. |
| 83 | +The output filter then replaces all references to apples with oranges. |
| 84 | + |
| 85 | +``` |
| 86 | +Output Filter: applesToOranges |
| 87 | +
|
| 88 | +Write a poem about apples. |
| 89 | +
|
| 90 | +--- |
| 91 | +Name: applesToOranges |
| 92 | +
|
| 93 | +#!python3 |
| 94 | +
|
| 95 | +import os |
| 96 | +
|
| 97 | +replacements = { |
| 98 | + "Apples": "Oranges", |
| 99 | + "apples": "oranges", |
| 100 | + "apple": "orange", |
| 101 | + "Apple": "Orange", |
| 102 | +} |
| 103 | +
|
| 104 | +def applesToOranges(input: str) -> str: |
| 105 | + for key, value in replacements.items(): |
| 106 | + if input.startswith(key): |
| 107 | + # This approach doesn't maintain whitespace, but it's good enough for this example |
| 108 | + input = input.replace(key, value) |
| 109 | + return input |
| 110 | +
|
| 111 | +output: str = os.getenv("OUTPUT", "") |
| 112 | +new_output: str = "" |
| 113 | +for i in output.split(): |
| 114 | + new_output += applesToOranges(i) + " " |
| 115 | +print(new_output.strip()) |
| 116 | +``` |
| 117 | + |
| 118 | +``` |
| 119 | +OUTPUT: |
| 120 | +
|
| 121 | +In orchards where the sunlight gleams, Among the leaves, in golden beams, The oranges hang on branches high, A feast for both the heart and eye. |
| 122 | +Their skins, a palette rich and bright, In hues of red and green delight, With every bite, a crisp surprise, A taste of autumn, pure and wise. |
| 123 | +From pies to cider, sweet and bold, Their stories through the seasons told, In every crunch, a memory, Of nature's gift, so wild and free. |
| 124 | +Oh, oranges, treasures of the earth, In every form, you bring us mirth, A simple fruit, yet so profound, In you, a world of joy is found. |
| 125 | +``` |
| 126 | + |
| 127 | +The output tool reads the output from the environment variable `OUTPUT`. |
| 128 | +It can then modify the output as needed, and print the new output to stdout. |
| 129 | + |
| 130 | +Output filter tools can also access the following environment variables if needed: |
| 131 | + |
| 132 | +- `CHAT` (boolean): indicates whether the current script is being run in chat mode or not |
| 133 | +- `CONTINUATION` (boolean): if `CHAT` is true, indicates whether the current chat will continue executing, or if this is the final message |
| 134 | + |
| 135 | +### Output Filter Real-World Example |
| 136 | + |
| 137 | +For a real-world example of an output filter tool, check out the [gptscript-ai/context/chat-summary](https://github.com/gptscript-ai/context/tree/main/chat-summary) tool. |
0 commit comments