Upload 3 files

app.py

@@ -26,6 +26,7 @@ import dashboard as DB
 import plans as P
 import time
 import site_pages as SP
+import report_flags as RF
 from humanizer import BRAND, RAINBOW, LOGO_SRC, FAVICON_SRC, FAVICON_PATH, MODES, OUT_DIR
 
 APP_NAME = os.getenv("APP_NAME", "CS AI Writing Suite")
@@ -258,8 +259,19 @@ def _smart_result(res, hi, mod, text_mode, rep=None, name=""):
     return tuple(out)
 
 
-def _smart_text_core(text, target, passes, temp, mode, hi, mod, full_ai, vocab, rep):
-    H.FULL_AI.set(bool(full_ai)); H.VOCAB_TARGET.set(float(vocab or 0))
+def _use_report(report):
+    """Uploaded AI / similarity report -> only the sentences it flags are rewritten (None = normal run)."""
+    if not report:
+        RF.ACTIVE.set(None)
+        return
+    try:
+        RF.ACTIVE.set(RF.load(report))
+    except RF.ReportError as err:
+        raise gr.Error(str(err))
+
+
+def _smart_text_core(text, target, passes, temp, mode, hi, mod, full_ai, vocab, report, rep):
+    H.FULL_AI.set(bool(full_ai)); H.VOCAB_TARGET.set(float(vocab or 0)); _use_report(report)
     text = (text or "").strip()
     if not text:
         raise gr.Error("Please paste some text.")
@@ -267,23 +279,23 @@ def _smart_text_core(text, target, passes, temp, mode, hi, mod, full_ai, vocab,
     return _smart_result(res, hi, mod, True, rep, "Pasted text")
 
 
-def _smart_docx_core(path, target, passes, temp, mode, hi, mod, full_ai, vocab, rep):
-    H.FULL_AI.set(bool(full_ai)); H.VOCAB_TARGET.set(float(vocab or 0))
+def _smart_docx_core(path, target, passes, temp, mode, hi, mod, full_ai, vocab, report, rep):
+    H.FULL_AI.set(bool(full_ai)); H.VOCAB_TARGET.set(float(vocab or 0)); _use_report(report)
     if not path or not str(path).lower().endswith(".docx"):
         raise gr.Error("Please upload a Word (.docx) file.")
     res = S.smart_docx(path, target, passes, temp, mode, hi, mod, rep)
     return _smart_result(res, hi, mod, False, rep, os.path.basename(str(path)))
 
 
-def smart_text(text, target, passes, temp, mode, hi, mod, full_ai, vocab, user, request: gr.Request = None):
+def smart_text(text, target, passes, temp, mode, hi, mod, full_ai, vocab, report, user, request: gr.Request = None):
     passes = P.max_passes(user, passes)
-    yield from _stream(_smart_text_core, (text, target, passes, temp, mode, hi, mod, bool(full_ai), vocab), 6, 0, user, request=request,
+    yield from _stream(_smart_text_core, (text, target, passes, temp, mode, hi, mod, bool(full_ai), vocab, report), 6, 0, user, request=request,
                        tool="Smart Humanize", paste=text)
 
 
-def smart_docx(path, target, passes, temp, mode, hi, mod, full_ai, vocab, user, request: gr.Request = None):
+def smart_docx(path, target, passes, temp, mode, hi, mod, full_ai, vocab, report, user, request: gr.Request = None):
     passes = P.max_passes(user, passes)
-    yield from _stream(_smart_docx_core, (path, target, passes, temp, mode, hi, mod, bool(full_ai), vocab), 5, 0, user, True, request,
+    yield from _stream(_smart_docx_core, (path, target, passes, temp, mode, hi, mod, bool(full_ai), vocab, report), 5, 0, user, True, request,
                        tool="Smart Humanize", words=P.doc_words(path) if path else None)
 
 
@@ -1222,6 +1234,11 @@ with gr.Blocks(title=f"{APP_NAME} · CS Software Solutions", theme=theme, css=CS
                                       "and checked by Groq. Slower. "
                                       "Leave it off for your own writing or lightly edited AI text.",
                                  elem_classes="cs-fullai")
+                s_report = gr.File(label="📑 Have an AI or similarity report? Upload it (optional)", file_types=[".pdf", ".docx"],
+                                   type="filepath", elem_classes="cs-report")
+                gr.HTML('<div class="cs-gauge-cap" style="text-align:left;margin:-4px 2px 4px">Turnitin, iThenticate, GPTZero, '
+                        'ZeroGPT, Copyleaks or any other checker: upload the report PDF as downloaded. Only the sentences '
+                        'the report highlights are rewritten; everything else stays exactly as it is.</div>')
                 s_hi, s_mod = thresholds()
             with gr.Tabs(elem_classes="sub-tabs"):
                 with gr.Tab("✍️ Paste text · free"):
@@ -1685,9 +1702,9 @@ with gr.Blocks(title=f"{APP_NAME} · CS Software Solutions", theme=theme, css=CS
     for _cb, _sl in ((s_full, s_vocab), (hp_full, hp_vocab), (hd_full, hd_vocab)):
         _cb.change(lambda v: gr.update(value=70 if v else 0), _cb, _sl, queue=False, **HID)
     sp_in.change(count_words_for, [sp_in, user_state], sp_count, queue=False, **HID)
-    sp_run = sp_btn.click(lambda: WAIT, None, sp_status, queue=False, **HID).then(smart_text, [sp_in, s_target, s_passes, s_temp, s_mode, s_hi, s_mod, s_full, s_vocab, user_state],
+    sp_run = sp_btn.click(lambda: WAIT, None, sp_status, queue=False, **HID).then(smart_text, [sp_in, s_target, s_passes, s_temp, s_mode, s_hi, s_mod, s_full, s_vocab, s_report, user_state],
                  [sp_status, sp_out, sp_prev, sp_scores, sp_file, sp_dl], **JOB(sp_status))
-    sd_run = sd_btn.click(lambda: WAIT, None, sd_status, queue=False, **HID).then(smart_docx, [sd_in, s_target, s_passes, s_temp, s_mode, s_hi, s_mod, s_full, s_vocab, user_state],
+    sd_run = sd_btn.click(lambda: WAIT, None, sd_status, queue=False, **HID).then(smart_docx, [sd_in, s_target, s_passes, s_temp, s_mode, s_hi, s_mod, s_full, s_vocab, s_report, user_state],
                  [sd_status, sd_prev, sd_scores, sd_file, sd_dl], **JOB(sd_status))
 
     # detector

report_flags.py

@@ -0,0 +1,175 @@
+"""
+Report-guided rewriting: read an AI-detection or similarity report (Turnitin, iThenticate, GPTZero, ZeroGPT,
+Copyleaks, our own CS report, ...) and find the passages it FLAGS, whatever the tool. A passage counts as flagged
+when the report marks it in any of the usual ways:
+
+  * highlight / underline / squiggly annotations (PDF)
+  * coloured text (similarity reports colour matched text by source)
+  * coloured shading drawn behind the text (AI-writing reports shade flagged sentences)
+  * Word reports (.docx): highlighted, shaded or coloured runs
+
+The flagged words are then matched to the sentences of the customer's document by word overlap (line breaks,
+hyphenation and page furniture in the report do not matter). Smart Humanize then rewrites exactly those
+sentences and nothing else.
+"""
+import contextvars
+import re
+
+MIN_SHARE = 0.5          # a sentence is flagged when at least half of its word trigrams were marked in the report
+_WORD = re.compile(r"[A-Za-z0-9]+(?:['’][A-Za-z]+)?")
+
+# set per job by the app when a report is uploaded (None = no report: normal Smart Humanize)
+ACTIVE = contextvars.ContextVar("report_flags", default=None)
+
+
+class ReportError(Exception):
+    """The report could not be used (no flagged passages found, scanned image, wrong file)."""
+
+
+def _norm_words(text):
+    text = re.sub(r"(\w)-\s*\n\s*(\w)", r"\1\2", text or "")      # de-hyphenate line breaks
+    return [w.lower().replace("’", "'") for w in _WORD.findall(text)]
+
+
+def _trigrams(words):
+    return {tuple(words[i:i + 3]) for i in range(len(words) - 2)}
+
+
+def _colourful(rgb, min_sat=0.12):
+    """True for a real colour (not black, white or grey)."""
+    if not rgb:
+        return False
+    r, g, b = rgb[:3]
+    return max(r, g, b) - min(r, g, b) >= min_sat
+
+
+def _int_rgb(c):
+    return ((c >> 16) & 255) / 255, ((c >> 8) & 255) / 255, (c & 255) / 255
+
+
+def _pdf_runs(path):
+    """Runs of flagged words from a PDF report (each run = consecutive marked words, in reading order)."""
+    import fitz  # PyMuPDF
+    try:
+        pdf = fitz.open(path)
+    except Exception as err:
+        raise ReportError(f"Could not open the report: {err}")
+    runs, n_words = [], 0
+    for page in pdf:
+        words = page.get_text("words")                 # x0, y0, x1, y1, text, block, line, word_no
+        n_words += len(words)
+        if not words:
+            continue
+        area = page.rect.width * page.rect.height
+        marks = []                                     # rectangles that mark text
+        for a in page.annots() or []:
+            if a.type[0] in (8, 9, 10):                # highlight, underline, squiggly
+                v = a.vertices or []
+                if len(v) >= 4:
+                    for i in range(0, len(v) - 3, 4):
+                        marks.append(fitz.Quad(v[i:i + 4]).rect)
+                else:
+                    marks.append(a.rect)
+        for d in page.get_drawings():                  # shading drawn behind flagged text
+            fill = d.get("fill")
+            r = d.get("rect")
+            if fill and r is not None and _colourful(fill, 0.08) and r.width * r.height < 0.5 * area \
+                    and r.height < 60:
+                marks.append(r)
+        coloured = set()                               # coloured text spans (similarity reports)
+        for blk in page.get_text("dict").get("blocks", []):
+            for line in blk.get("lines", []):
+                for sp in line.get("spans", []):
+                    if sp.get("text", "").strip() and _colourful(_int_rgb(sp.get("color", 0))):
+                        coloured.add(tuple(round(x, 1) for x in sp["bbox"]))
+        col_rects = [fitz.Rect(b) for b in coloured]
+
+        def marked(w):
+            cx, cy = (w[0] + w[2]) / 2, (w[1] + w[3]) / 2
+            p = fitz.Point(cx, cy)
+            return any(r.contains(p) for r in marks) or any(r.contains(p) for r in col_rects)
+
+        cur = []
+        for w in sorted(words, key=lambda w: (w[5], w[6], w[7])):
+            if marked(w):
+                cur.append(w[4])
+            elif cur:
+                runs.append(" ".join(cur))
+                cur = []
+        if cur:
+            runs.append(" ".join(cur))
+    if n_words == 0:
+        raise ReportError("This report has no readable text (it looks like a scan or image). Download the report "
+                          "as a PDF from the checker's website and upload that file.")
+    return runs
+
+
+def _docx_runs(path):
+    """Runs of flagged text from a Word report: highlighted, shaded or coloured runs."""
+    from docx import Document
+    from docx.oxml.ns import qn
+    runs = []
+    for p in Document(path).paragraphs:
+        cur = []
+        for r in p.runs:
+            rpr = r._r.find(qn("w:rPr"))
+            flag = False
+            if rpr is not None:
+                hl = rpr.find(qn("w:highlight"))
+                shd = rpr.find(qn("w:shd"))
+                col = rpr.find(qn("w:color"))
+                flag = (hl is not None and hl.get(qn("w:val")) not in (None, "none")) or \
+                       (shd is not None and (shd.get(qn("w:fill")) or "auto").lower() not in ("auto", "ffffff")) or \
+                       (col is not None and _colourful(_hex_rgb(col.get(qn("w:val")))))
+            if flag and r.text.strip():
+                cur.append(r.text)
+            elif cur:
+                runs.append(" ".join(cur))
+                cur = []
+        if cur:
+            runs.append(" ".join(cur))
+    return runs
+
+
+def _hex_rgb(h):
+    try:
+        h = (h or "").strip()
+        return int(h[0:2], 16) / 255, int(h[2:4], 16) / 255, int(h[4:6], 16) / 255
+    except Exception:
+        return None
+
+
+class Flags:
+    """The flagged passages of one report; .match(sentence) says whether a document sentence was flagged."""
+
+    def __init__(self, runs):
+        self.runs = [r for r in runs if len(_norm_words(r)) >= 3]
+        self.tri = set()
+        for r in self.runs:
+            self.tri |= _trigrams(_norm_words(r))
+        self.matched = 0
+
+    def match(self, sentence):
+        t = _trigrams(_norm_words(sentence))
+        if not t:
+            return False
+        hit = sum(1 for x in t if x in self.tri) / len(t) >= MIN_SHARE
+        self.matched += hit
+        return hit
+
+
+def load(path):
+    """Read a report file -> Flags. Raises ReportError with a customer-friendly message."""
+    p = str(path or "")
+    low = p.lower()
+    if low.endswith(".pdf"):
+        runs = _pdf_runs(p)
+    elif low.endswith(".docx"):
+        runs = _docx_runs(p)
+    else:
+        raise ReportError("Please upload the report as a PDF (or a Word .docx with the flagged text highlighted).")
+    flags = Flags(runs)
+    if not flags.tri:
+        raise ReportError("No highlighted or flagged passages were found in this report. Upload the full report PDF "
+                          "as downloaded from the checker (the version with the coloured / highlighted text).")
+    return flags

smart.py

@@ -19,6 +19,7 @@ import detector as D
 import textparse as TP
 import docx_report as DX
 import humanizer as H
+import report_flags as R
 
 W_T = qn("w:t")
 _XML_SPACE = "{http://www.w3.org/XML/1998/namespace}space"
@@ -191,6 +192,12 @@ def _target_change():
 
 
 def _done_early(k, pct, target, chg):
+    if R.ACTIVE.get() is not None:                     # report-guided: always rewrite what the report flagged
+        return False
+    return _done_early_inner(k, pct, target, chg)
+
+
+def _done_early_inner(k, pct, target, chg):
     """Normal mode: stop when the AI score is 0 or (from pass 2) at/below the target.
     100% AI mode: at least one rewrite, and the new-wording target must also be reached."""
     if not _all_sentences():
@@ -204,6 +211,8 @@ def _better(cand, best, k):
     then the lowest AI score; below the target, the version with the most new wording wins."""
     if best is None:
         return True
+    if R.ACTIVE.get() is not None:                     # report-guided: the rewritten version is the result
+        return k > 1
     if not _all_sentences():
         return cand[0] < best[0] or (k > 1 and cand[0] == best[0])
     if best[1] is None:                                # best is still the untouched original
@@ -275,6 +284,9 @@ def smart_docx(path, target, passes, temperature, mode, hi, mod, rep):
     log, first, note = [], None, ""
     orig_text = "\n".join(b.text for b in D.docx_blocks(doc) if b.kind == "para")
     TP.GLOSSARY.set(TP.build_glossary([orig_text]))     # the document's key terms stay untouched
+    rpt = R.ACTIVE.get()                               # uploaded AI / similarity report: rewrite only what it flags
+    if rpt is not None:
+        passes = 1
     best = None                                        # (pct, docx bytes, scored blocks)
     step = 0.8 / max(1, passes)
     for k in range(1, int(passes) + 2):                # last iteration = final check only
@@ -299,12 +311,18 @@ def smart_docx(path, target, passes, temperature, mode, hi, mod, rep):
             if b.kind != "para" or b.el is None:
                 continue
             for sp, s, p in zip(b.cspans, b.sents, b.scores):
-                if sp and (full or p >= mod) and len(s.split()) >= 5:
+                pick = rpt.match(s) if rpt is not None else (full or p >= mod)
+                if sp and pick and len(s.split()) >= 5:
                     sub = safe_subspan(b.el, *sp)
                     if sub:
                         jobs.append((b, sub, b.raw[sub[0]:sub[1]]))
                     else:
                         locked += 1
+        if not jobs and rpt is not None:
+            raise gr.Error("None of the sentences in your document match the passages flagged in the report. "
+                           "Check that the report belongs to this document.")
+        if rpt is not None:
+            log.append(f"Report: {len(jobs) + locked} of your sentences match passages the report flagged")
         if not jobs:
             log.append(_log_line(k, stats["pct"]) + f" — no rewritable sentences left ({locked} locked)")
             break
@@ -359,6 +377,9 @@ def smart_text(text, target, passes, temperature, mode, hi, mod, rep):
     paras = [p for p, _ in parts]
     orig = TP.join_paste(parts)
     log, first, note, best = [], None, "", None       # best = (pct, paras, blocks)
+    rpt = R.ACTIVE.get()                               # uploaded AI / similarity report: rewrite only what it flags
+    if rpt is not None:
+        passes = 1
     step = 0.8 / max(1, passes)
     for k in range(1, int(passes) + 2):
         base = 0.05 + step * (k - 1)
@@ -377,7 +398,13 @@ def smart_text(text, target, passes, temperature, mode, hi, mod, rep):
             break
         full = _all_sentences()                        # 100% AI box / vocabulary level: every sentence
         jobs = [(bi, si, s) for bi, b in enumerate(blocks) if b.kind == "para"
-                for si, (s, p) in enumerate(zip(b.sents, b.scores)) if (full or p >= mod) and len(s.split()) >= 5]
+                for si, (s, p) in enumerate(zip(b.sents, b.scores))
+                if (rpt.match(s) if rpt is not None else (full or p >= mod)) and len(s.split()) >= 5]
+        if rpt is not None:
+            if not jobs:
+                raise gr.Error("None of your sentences match the passages flagged in the report. "
+                               "Check that the report belongs to this text.")
+            log.append(f"Report: {len(jobs)} of your sentences match passages the report flagged")
         if not jobs:
             log.append(_log_line(k, stats["pct"]) + " — no rewritable sentences left")
             break