Portkey enhances this capability with the raw prompt feature of prompt templates. You can easily add few-shot learning examples to your templates with raw prompt and dynamically update them whenever you want, without needing to modify the prompt templates!

How does it work?

Let’s consider a use case where, given a candidate profile and a job description, the LLM is expected to output candidate notes in a specific JSON format.

This is how our raw prompt looks:

[
    {
        "role": "system",
        "message": "You output candidate notes in JSON format when given a candidate profile and a job description.",
    },

    {{few_shot_examples}},

    {
        "role": "user",
        "message": "Candidate Profile: {{profile}} \n Job Description: {{jd}}"
    },
]

Let’s define our variables:

As you can see, we have added variables few_shot_examples, profile, and jd in the above examples.

profile = "An experienced data scientist with a PhD in Computer Science and 5 years of experience working with machine learning models in the healthcare industry."

jd = "We are seeking a seasoned data scientist with a strong background in machine learning, ideally with experience in the healthcare sector. The ideal candidate should have a PhD or Master's degree in a relevant field and a minimum of 5 years of industry experience."

And now let’s add some examples with the expected JSON structure:

few_shot_examples =

[

 {

  "role": "user",

  "content": "Candidate Profile: Experienced software engineer with a background in developing scalable web applications using Python. Job Description: We’re looking for a Python developer to help us build and scale our web platform.",

 },

 {

  "role": "assistant",

  "content": "{'one-line-intro': 'Experienced Python developer with a track record of building scalable web applications.', 'move-forward': 'Yes', 'priority': 'P1', 'pros': '1. Relevant experience in Python. 2. Has built and scaled web applications. 3. Likely to fit well with the job requirements.', 'cons': 'None apparent from the provided profile.'}",

 },

 {

  "role": "user",

  "content": "Candidate Profile: Recent graduate with a degree in computer science and a focus on data analysis. Job Description: Seeking a seasoned data scientist to analyze large data sets and derive insights."

 },

 {

  "role": "assistant",

  "content": "{'one-line-intro': 'Recent computer science graduate with a focus on data analysis.', 'move-forward': 'Maybe', 'priority': 'P2', 'pros': '1. Has a strong educational background in computer science. 2. Specialized focus on data analysis.', 'cons': '1. Lack of professional experience. 2. Job requires a seasoned data scientist.' }"

  }

]

In this configuration, {{few_shot_examples}} is a placeholder for the few-shot learning examples, which are dynamically provided and can be updated as needed. This allows the LLM to adapt its responses to the provided examples, facilitating versatile and context-aware outputs.

Putting it all together in Portkey’s prompt manager:

  1. Go to the “Prompts” page on https://app.portkey.ai/ and Create a new Prompt template with your preferred AI provider.
  2. Selecting Chat mode will enable the Raw Prompt feature:
  1. Click on it and paste the raw prompt code from above. And that’s it! You have your dynamically updatable few shot prompt template ready to deploy.

Deploying the Prompt with Portkey

Deploying your prompt template to an API is extremely easy with Portkey. You can use our Prompt Completions API to use the prompt we created.

from portkey_ai import Portkey

client = Portkey(
    api_key="PORTKEY_API_KEY",  # defaults to os.environ.get("PORTKEY_API_KEY")
)

prompt_completion = client.prompts.completions.create(
    prompt_id="Your Prompt ID", # Add the prompt ID we just created
    variables={
       few_shot_examples: fseObj,
       profile: "",
       jd: ""
    }
)

print(prompt_completion)
# We can also override the hyperparameters
prompt_completion = client.prompts.completions.create(
    prompt_id="Your Prompt ID", # Add the prompt ID we just created
    variables={
       few_shot_examples: fseObj,
       profile: "",
       jd: ""
    }
    max_tokens=250,
    presence_penalty=0.2
)
print(prompt_completion)

You can pass your dynamic few shot learning examples with the few_shot_examples variable, and start using the prompt template in production!

Detailed Guide on Few-Shot Prompting

We recommend this guide detailing the research as well as edge cases for few-shot prompting.

Support

Facing an issue? Reach out on [email protected] for a quick resolution.