Closed Domain Q&A

TL;DR

• The core of Closed-domain Q&A: answers must be strictly limited to the given facts/context. No "freestyling."
• The most common problem: the model fills in information that doesn't exist (hallucination), especially in specialized domains (medical/legal/finance).
• Production tip: enforce "use exclusively the information above," and pair it with evaluation (cross-check every claim).

Background

The following prompt tests an LLM's capabilities to answer closed-domain questions which involves answering questions within a specific topic or domain.

Note: due to the challenging nature of the task, LLMs are likely to hallucinate when they have no knowledge regarding the question.

How to Apply

This example transfers directly to "rewrite/generate documents from given data" scenarios:

• Input is facts (bullet list / table / JSON)
• Output is a structured note (medical note, case summary, policy memo, etc.)
• Require "use only given information" to avoid speculative content

How to Iterate

1. Define an explicit output template (sections + fields) to reduce the model's room to improvise
2. Require citations: each output paragraph tags the corresponding fact line number/field name (if input can be numbered)
3. Add self-check: before output, list a "claims list," then cross-reference each claim against the input
4. Add "unknown policy": when information is missing, output "Unknown" or ask clarifying questions

Self-check Rubric

• Does the output contain specific information not in the input (diagnosis, BMI, complications, treatment recommendations, etc.)?
• Are facts and speculation mixed together? Can assumptions be clearly labeled?
• Does it follow "exclusively the information above"?

Practice

Exercise: use your own work materials for a closed-domain rewrite.

1. Prepare a set of facts (10-20 items, ideally numbered)
2. Have the model output a structured note
3. Then use a truthfulness rubric for claim-by-claim review (can pair with /prompt-truthfulness-identify-hallucination)

Prompt

Patient's facts:

-   20 year old female
-   with a history of anerxia nervosa and depression
-   blood pressure 100/50, pulse 50, height 5'5''
-   referred by her nutrionist but is in denial of her illness
-   reports eating fine but is severely underweight

Please rewrite the data above into a medical note, using exclusively the information above.

Code / API

OpenAI (Python)

from openai import OpenAI

client = OpenAI()

response = client.chat.completions.create(
    model="gpt-4",
    messages=[
        {
            "role": "user",
            "content": "Patient's facts:\n- 20 year old female\n- with a history of anerxia nervosa and depression\n- blood pressure 100/50, pulse 50, height 5'5''\n- referred by her nutrionist but is in denial of her illness\n- reports eating fine but is severely underweight\n\nPlease rewrite the data above into a medical note, using exclusively the information above.",
        }
    ],
    temperature=1,
    max_tokens=500,
    top_p=1,
    frequency_penalty=0,
    presence_penalty=0,
)

Fireworks (Python)

import fireworks.client

fireworks.client.api_key = "<FIREWORKS_API_KEY>"

completion = fireworks.client.ChatCompletion.create(
    model="accounts/fireworks/models/mixtral-8x7b-instruct",
    messages=[
        {
            "role": "user",
            "content": "Patient's facts:\n- 20 year old female\n- with a history of anerxia nervosa and depression\n- blood pressure 100/50, pulse 50, height 5'5''\n- referred by her nutrionist but is in denial of her illness\n- reports eating fine but is severely underweight\n\nPlease rewrite the data above into a medical note, using exclusively the information above.",
        }
    ],
    stop=["<|im_start|>", "<|im_end|>", "<|endoftext|>"],
    stream=True,
    n=1,
    top_p=1,
    top_k=40,
    presence_penalty=0,
    frequency_penalty=0,
    prompt_truncate_len=1024,
    context_length_exceeded_behavior="truncate",
    temperature=0.9,
    max_tokens=4000,
)

Reference

• Sparks of Artificial General Intelligence: Early experiments with GPT-4 (13 April 2023)