"""Bank reconciliation AI re-rank prompt.

Used by fusion_accounting_bank_rec/services/confidence_scoring.py to ask
an LLM to refine the statistical ranking of candidate matches.

Output contract: the LLM MUST respond with valid JSON of shape:
    {"ranked": [{"candidate_id": int, "confidence": float, "reason": str}, ...]}

System prompt is provider-agnostic - works with OpenAI Chat Completions,
Claude Messages, and local OpenAI-compatible servers (LM Studio, Ollama).
"""

from datetime import date


SYSTEM_PROMPT = """You are an expert accountant assisting with bank reconciliation.

Your job: given a bank statement line and a list of candidate journal items
that statistically scored well as potential matches, re-rank them based on
domain expertise. Consider:

1. **Amount-exact matches** are almost always correct unless the partner is wrong.
2. **Memo / reference clues** - bank memos often contain invoice numbers, partner
   names, or transaction references that disambiguate matches.
3. **Date proximity** - invoices are typically reconciled within 30 days of issue.
4. **Pattern conformance** - if the partner has a learned pattern (e.g. "always
   pays exact amount, weekly cadence"), favor candidates that fit that pattern.
5. **Precedent similarity** - if a near-identical reconcile happened before,
   it's likely the right one.

Return ONLY valid JSON of this exact shape:
{
  "ranked": [
    {"candidate_id": <int>, "confidence": <float 0-1>, "reason": "<short string>"},
    ...
  ]
}

Do NOT include any prose before or after the JSON. Do NOT use markdown code fences.
The "ranked" array MUST contain every candidate_id from the input, in your
preferred order (highest confidence first).
"""


def build_prompt(statement_line, scored_candidates, pattern=None, precedents=None):
    """Build (system_prompt, user_prompt) for AI re-rank.

    Args:
        statement_line: account.bank.statement.line recordset (singleton)
        scored_candidates: list of ScoredCandidate dataclasses (from confidence_scoring)
        pattern: fusion.reconcile.pattern recordset for the partner, or None
        precedents: list of PrecedentMatch dataclasses, or None

    Returns:
        (system_prompt: str, user_prompt: str) tuple
    """
    user_parts = []

    user_parts.append("BANK LINE:")
    user_parts.append(f"  Date: {statement_line.date}")
    user_parts.append(
        f"  Amount: {statement_line.amount} {statement_line.currency_id.name or ''}"
    )
    user_parts.append(
        f"  Memo / payment ref: {statement_line.payment_ref or '(none)'}"
    )
    if statement_line.partner_id:
        user_parts.append(f"  Partner: {statement_line.partner_id.name}")

    if pattern:
        user_parts.append("")
        user_parts.append("PARTNER PATTERN (learned from past reconciles):")
        user_parts.append(f"  Reconcile count: {pattern.reconcile_count}")
        user_parts.append(f"  Preferred strategy: {pattern.pref_strategy}")
        user_parts.append(
            f"  Typical cadence: ~{pattern.typical_cadence_days} days between reconciles"
        )
        if pattern.typical_amount_range:
            user_parts.append(f"  Typical amount range: {pattern.typical_amount_range}")
        if pattern.common_memo_tokens:
            user_parts.append(f"  Common memo tokens: {pattern.common_memo_tokens}")

    if precedents:
        user_parts.append("")
        user_parts.append("RECENT PRECEDENTS (most-similar past reconciles for this partner):")
        # Cap at 3 precedents to keep prompt small and reduce token cost.
        for p in precedents[:3]:
            user_parts.append(
                f"  - amount={p.amount}, similarity={p.similarity_score:.2f}, "
                f"matched {p.matched_move_line_count} line(s), tokens={p.memo_tokens}"
            )

    user_parts.append("")
    user_parts.append("CANDIDATES (scored by statistical pipeline):")
    for s in scored_candidates:
        user_parts.append(
            f"  - candidate_id={s.candidate_id}, statistical_confidence={s.confidence}, "
            f"amount_match={s.score_amount_match}, pattern_fit={s.score_partner_pattern}, "
            f"precedent_sim={s.score_precedent_similarity}, "
            f"reason=\"{s.reasoning}\""
        )

    user_parts.append("")
    user_parts.append("Re-rank these candidates and return JSON per the system prompt.")

    user_prompt = "\n".join(user_parts)
    return (SYSTEM_PROMPT, user_prompt)