#!/bin/bash
# One-shot commit for the v2.0 Hebrew RAG production milestone.
# 24/24 regression tests pass.
set -e
cd "$( dirname "${BASH_SOURCE[0]}" )/.."

echo "▸ Staging code changes..."
git add \
    pipeline.py \
    api/fastapi_app.py \
    generate/extractive.py \
    middleware/adversarial_prompt.py \
    static/index.html \
    verify/citations.py \
    verify/structural_nli.py \
    scripts/deploy_hebrew_rag.sh \
    scripts/test_hebrew_rag.sh \
    scripts/commit_v2.sh \
    HEBREW_RAG_RUNBOOK.md \
    PRODUCTION_GAP_ANALYSIS.md

echo "▸ Current diff summary:"
git diff --cached --stat

echo ""
echo "▸ Creating commit..."
git commit -m "$(cat <<'MSG'
Hebrew RAG v2.0 — 24/24 regression tests pass

Pipeline (retrieval correctness):
- Symmetric language filter: Hebrew→Hebrew-only, English→drops Hebrew,
  empty-fallback returns no results (no more language leakage into EN queries).
- "סעיף N" intent boost (+5.0 for statute, -0.5 for caselaw) + section
  sidecar that injects the actual §N statute from _indexed_docs when
  BM25 buries it behind shorter noisy chunks.
- key_terms exact-match boost for Hebrew topical queries (gated to short
  queries without caselaw intent markers).
- Global post-fusion filter pass + citation sidecar: honors query.filters
  across ALL retrievers (dense/gematria/hilbert don't natively filter).
- "הלכת X" landmark sidecar: exact key_terms match → force-injection
  from supplement, guards landmark rulings against bulk caselaw
  referencing them by name.

Data:
- heb_law_supplement.jsonl: added [הלכת X] markers + key_terms for all
  5 landmark rulings (apropim, bait_yules, kal_binyan, klinor, mivtahim).

API:
- /v1/suggestions (public, read-only) — frontend no longer hits the
  admin-gated /v1/admin/query_stats, eliminates 401 console noise.

Frontend:
- Chat history (localStorage, up to 50 conversations, search, replay)
- Export to Markdown / Word (.doc HTML) / HTML / PDF (print) / Clipboard
- Header buttons: 🕒 היסטוריה, ＋ חדש, 📤 ייצוא

Ops scripts:
- deploy_hebrew_rag.sh restart → auto-detects empty index and
  re-uploads JSONL automatically (no more stale-index confusion)
- test_hebrew_rag.sh: robust JSON-body builder via python3 json.dumps,
  no more set -e + quote-escaping silent failures. 24 tests total.

Verifiers:
- Skip disclaimer sentences from citation coverage / structural NLI
  scoring (avoid false-positive unsupported_claim alerts).

Adversarial prompt detection:
- 7 new Hebrew patterns (SEVERITY_CRITICAL) for "שכח/התעלם/תתעלם"
  + SYSTEM:/USER:/ASSISTANT: prefixes + mixed EN-HE injection chains.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
MSG
)"

echo ""
echo "▸ Tagging..."
git tag -a heb-rag-v2.0 -m "Hebrew legal RAG — 24/24 regression tests, production-ready retrieval"

echo ""
echo "✅ Done. Commit + tag created."
echo "   git log -1 --oneline"
git log -1 --oneline
echo ""
echo "   git tag -l 'heb-rag-*'"
git tag -l 'heb-rag-*'