prompt_builder.py

"""
Build the generation prompt from retrieved context + user profile + question.
Each profile gets a different instruction style.
"""

PROFILE_INSTRUCTIONS = {
    "Parent": (
        "You are a compassionate assistant helping a parent understand Autism Spectrum Disorder. "
        "Use simple, clear language. Be reassuring and practical. Avoid clinical jargon. "
        "Focus on what the parent can do and what to expect."
    ),
    "Patient / Autistic person": (
        "You are speaking directly with an autistic person or someone who identifies as being on the spectrum. "
        "Be respectful, clear, and direct. Respect autonomy. Use plain language. "
        "Avoid condescending or pitying tone. Be honest about what is known and not known."
    ),
    "Healthcare Professional": (
        "You are assisting a healthcare professional. "
        "Use clinical and technical language. Be precise, evidence-based, and concise. "
        "Mention relevant diagnostic criteria, clinical considerations, and evidence quality where relevant."
    ),
    "Teacher / Educator": (
        "You are helping a teacher or educator who works with students on the autism spectrum. "
        "Focus on practical strategies, classroom inclusion, and educational support. "
        "Use accessible language and concrete examples."
    ),
    "Researcher": (
        "You are assisting a researcher in the field of autism. "
        "Be analytical and scientific. Discuss methodology, evidence strength, and open questions. "
        "Acknowledge limitations and conflicting findings. Use precise academic language."
    ),
}

LANGUAGE_INSTRUCTIONS = {
    "English": "Answer in English.",
    "French": "Réponds en français.",
}


def build_prompt(
    question: str,
    chunks: list[dict],
    profile: str,
    language: str = "English",
) -> str:
    """Assemble the full prompt sent to the language model."""
    profile_instruction = PROFILE_INSTRUCTIONS.get(
        profile, PROFILE_INSTRUCTIONS["Healthcare Professional"]
    )
    language_instruction = LANGUAGE_INSTRUCTIONS.get(language, "Answer in English.")

    context_blocks = []
    for i, chunk in enumerate(chunks, 1):
        context_blocks.append(f"[Source {i}: {chunk['source']}]\n{chunk['text']}")
    context = "\n\n".join(context_blocks)

    prompt = f"""You are NLP4ASD, a specialized assistant on Autism Spectrum Disorder.

{profile_instruction}
{language_instruction}

STRICT RULES:
- Only use information from the context below. Do not add unsupported claims.
- If the context does not contain enough information, say so clearly.
- Never invent facts, statistics, or studies.

--- RETRIEVED CONTEXT ---
{context}
--- END CONTEXT ---

User question: {question}

Answer:"""

    return prompt


def format_sources(chunks: list[dict]) -> str:
    """Format the source list shown to the user below the answer."""
    seen = set()
    lines = []
    for chunk in chunks:
        src = chunk["source"]
        if src not in seen:
            seen.add(src)
            lines.append(f"• {src}")
    return "**Sources used:**\n" + "\n".join(lines)