Generating Knowledge Prompts

Two-step pattern: (1) elicit knowledge, (2) answer using that knowledge.

Python
def generate_response(prompt, model="gpt-4o-mini", temperature=0.7, max_tokens=500):
    from openai import OpenAI
    client = OpenAI()
    resp = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        temperature=temperature,
        max_tokens=max_tokens,
    )
    return resp.choices[0].message.content.strip()

question = "Explain quantum entanglement and its potential applications."

# Step 1: Generate knowledge
knowledge_prompt = f"""Generate relevant background knowledge needed to answer this question well.
List key facts, concepts, and context. Do not answer the question yet.

Question: {question}"""
knowledge = generate_response(knowledge_prompt, temperature=0.3)

# Step 2: Answer using the knowledge
answer_prompt = f"""Question: {question}

Relevant knowledge:
{knowledge}

Using only the knowledge above (and sound reasoning), answer the question clearly and accurately."""
answer = generate_response(answer_prompt, temperature=0.5)

Lower temperature (0.2–0.4) for knowledge generation favors factual recall over creativity; moderate temperature (0.5–0.7) for the final answer allows natural synthesis.

Progress:
- [ ] Step 1: Define the task and exact question
- [ ] Step 2: Write a knowledge-generation prompt (specific, scoped, format-constrained)
- [ ] Step 3: Generate knowledge (low temperature, check relevance/accuracy)
- [ ] Step 4: Build final prompt combining question + generated knowledge
- [ ] Step 5: Generate final answer
- [ ] Step 6: Validate — check for irrelevance, repetition, hallucination
- [ ] Step 7: Iterate on prompt wording if quality is poor

Step 1 — Define task/question. Be explicit and narrow. Vague questions produce vague knowledge.

Step 2 — Craft the knowledge prompt. Techniques that improve quality:

• Specify desired format (free text, bullet list, JSON schema).
• Give 1–2 examples of the kind of knowledge wanted (few-shot).
• Inject keywords/subtopics to steer coverage (e.g., "cover political, economic, and military factors").
• Explicitly say "do not answer the question yet" to prevent premature conflation of steps.

Step 3 — Generate knowledge. Use low temperature for factual stability. If using JSON output, parse with error handling — LLMs don't always emit valid JSON:

Python
import json, re

def parse_json_knowledge(raw_text):
    match = re.search(r"\[.*\]", raw_text, re.DOTALL)
    if not match:
        return None
    try:
        return json.loads(match.group(0))
    except json.JSONDecodeError:
        return None  # fall back to raw text or retry

Step 4 — Build final prompt. Concatenate question + generated knowledge + explicit instruction to use it ("using the knowledge above...").

Step 5 — Generate final answer.

Step 6 — Validate. Check: Is the knowledge actually relevant? Any obvious factual errors? Did the final answer ignore the knowledge and hallucinate anyway? Spot-check against trusted sources for high-stakes use cases.

Step 7 — Iterate. If knowledge is off-target, narrow the prompt scope, add keywords, or supply format constraints.

Example 1: Scientific explanation Input: "Explain the concept of quantum entanglement and its potential applications." Output (knowledge step, abridged): "Quantum entanglement: a phenomenon where two particles become correlated such that measuring one instantly affects the state of the other, regardless of distance. Key figures: Einstein ('spooky action'), Bell's theorem... Applications: quantum computing, quantum cryptography (QKD), quantum teleportation of information..." Output (final answer): A coherent explanation that references entanglement correlation, Bell's theorem context, and lists computing/cryptography applications — grounded in the generated facts rather than free-associated.

Example 2: Historical causes Input: "What were the main causes of World War I?" Knowledge-prompt addition: "First, generate relevant knowledge about the political and social climate in Europe leading up to World War I." Output: Knowledge covering alliance systems, militarism, nationalism, the assassination of Archduke Franz Ferdinand → final answer synthesizes these into a structured causal explanation (MAIN: militarism, alliances, imperialism, nationalism + the triggering event).

Example 3: Structured JSON knowledge Input: "What are the benefits of using renewable energy sources?" Knowledge-prompt: Request JSON array of {"source": ..., "benefits": ...} objects. Output: Parsed JSON list (solar, wind, hydro, geothermal with benefits) fed into final prompt → answer organized source-by-source rather than generic.