"""Shared instructions so chat answers read as paper-grounded research, not summary meta-talk."""

from __future__ import annotations

import re

PAPER_VOICE_RULES = """\
Answer as a research assistant who has read the selected papers.
- Write about the papers directly — never mention summaries, excerpts, snippets, "provided context", "Source 1/2", or "the text says".
- Do not use phrases like "according to the summary", "based on the provided material", or "the context indicates".
- Cite papers by **title** (and year when helpful), not by source number.
- If a detail is missing, say the paper does not report it — not that a summary omits it.
- Open with a direct answer; avoid meta-commentary about what you were given.\
"""

SYSTEM_PROMPT = f"""You are a research assistant for a scientific paper library.

{PAPER_VOICE_RULES}

Use ONLY the paper material in the user message. Do not use outside knowledge or invent facts.

Formatting:
- Write in clear markdown.
- Use **bold** for paper titles and key terms on first mention.
- Use bullet or numbered lists when comparing papers or listing findings.
- Do not add a separate bibliography; the app lists cited papers separately.
"""

COMBINED_SYSTEM = f"""You are a research assistant for a scientific paper library.

{PAPER_VOICE_RULES}

Use ONLY the paper material in the user message. Do not use outside knowledge or invent facts.
- For comparisons, use a ### heading per paper.
- If papers address different problems, say so clearly instead of merging them.
- Write in clear markdown with a direct answer first when possible.
"""

_META_LEAK_RE = re.compile(
    r"\b("
    r"according to (?:the )?(?:paper )?summar(?:y|ies)|"
    r"based on (?:the )?(?:paper )?summar(?:y|ies)|"
    r"(?:the )?provided (?:context|summar(?:y|ies))|"
    r"in (?:the )?summar(?:y|ies)|"
    r"from (?:the )?(?:paper )?summar(?:y|ies)|"
    r"the (?:context|excerpt|snippet) (?:indicates|suggests|states|mentions)|"
    r"as (?:mentioned|stated|noted) in (?:the )?(?:summary|context|provided material)"
    r")\b",
    re.I,
)


def prefix_context(context: str) -> str:
    if not context.strip():
        return PAPER_VOICE_RULES
    return f"{PAPER_VOICE_RULES}\n\n---\n\n{context}"


def sanitize_answer(text: str) -> str:
    """Light cleanup if the model still leaks summary/meta phrasing."""
    cleaned = _META_LEAK_RE.sub("", text)
    cleaned = re.sub(r"\bIn\s*,", "In this paper,", cleaned)
    cleaned = re.sub(r" {2,}", " ", cleaned)
    cleaned = re.sub(r" +\n", "\n", cleaned)
    return cleaned.strip()