ok so here's a very basic fewshot, it works fine: ``` import openai, transformers client = openai.OpenAI(api_key="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", base_url='https://openrouter.ai/api/v1') def complete(model, prompt, end): completion = client.completions.create(model=model, prompt=prompt, stop=end, temperature=0.0) return completion.choices[0].text class PrefixInfixSuffixAction: template = 'Follow the examples:\n{left} => {right}\n' model = 'deepseek-ai/DeepSeek-V3' def run(self, pairs, example, template=template, model=model): prefix, subtemplate = template.split('{left}',1) infix, suffix = subtemplate.split('{right}',1) prompt = prefix + suffix.join([str(left) + infix + str(right) for left, right in pairs]) + str(example) + infix return complete(model, prompt, suffix) print(PrefixInfixSuffixAction().run([["1+1",2]],"9+7")) # outputs 16 ``` now, ummmmmm say i want to make a prompt that does singleshot .... 1935 ... the output is the structure used to perform the action; that would be the prompt. so for example, "template" here is the prompt. the input is the behavior of the action. so if "9+7 => 16\n" were the _input_, then "Follow the examples:\n1+1 => 2\n" would be the _output_, the second half of the pair. Something that's missing here is: - storage of data - judging if data is good or bad, right or wrong