| """The proofreader's prompt protocol: the exact text the model was trained on. |
| |
| The wordspace app sends two chat messages to the model, a system prompt and |
| the user's text, at temperature 0, and keeps only the "style" suggestions |
| whose "original" appears verbatim in the text (spelling and grammar come |
| from the operating system's checker). The model answers with one JSON object |
| and stops at its end-of-text token. |
| |
| The wording must stay identical, character for character: the model has |
| only ever seen this text. |
| """ |
| import json |
|
|
| SYSTEM = """You are a proofreader. You reply ONLY with JSON. |
| |
| Find mistakes in the user's text. For each one output an object with: |
| - "type": exactly one of "spelling", "grammar", "style" |
| - "original": the exact substring from the text, copied verbatim |
| - "replacement": the corrected substring |
| |
| Rules, in order of importance: |
| 1. MINIMAL SPAN. "original" must be the SHORTEST substring that has to |
| change — usually ONE word. A misspelled word gives original="thatword" |
| only. NEVER quote a whole sentence when a word is enough. |
| 2. "original" MUST appear verbatim in the text, character for character. |
| Never paraphrase it, never fix it inside "original". |
| 3. One mistake = one object. Do not merge several mistakes into one. |
| 4. "spelling" = the word does not exist. "grammar" = wrong agreement, tense |
| or preposition. "style" = correct but wordy or clumsy. |
| 5. Write in the SAME language as the text. Never translate. |
| 6. Be conservative: if the text is fine, return an empty list. |
| |
| Answer with this exact shape: |
| {"suggestions": [{"type": "...", "original": "...", "replacement": "..."}]}""" |
|
|
| |
| |
| LANGUAGE_NAMES = {"it": "Italian", "en": "English", "es": "Spanish", "fr": "French", |
| "de": "German", "pt": "Portuguese"} |
|
|
|
|
| def user(text, language=None): |
| """The user message: "Text (Italian):\\n..." or "Text:\\n...".""" |
| message = "Text" |
| if language: |
| message += " (%s)" % language |
| return message + ":\n" + text |
|
|
|
|
| def messages(text, language=None): |
| """The two chat messages, for `tokenizer.apply_chat_template`.""" |
| return [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user(text, language)}] |
|
|
|
|
| def first_json_object(answer): |
| """The first JSON object with balanced braces, as the app reads it.""" |
| start = answer.find("{") |
| while start >= 0: |
| depth, in_string, escape = 0, False, False |
| for i in range(start, len(answer)): |
| ch = answer[i] |
| if in_string: |
| if escape: |
| escape = False |
| elif ch == "\\": |
| escape = True |
| elif ch == '"': |
| in_string = False |
| elif ch == '"': |
| in_string = True |
| elif ch == "{": |
| depth += 1 |
| elif ch == "}": |
| depth -= 1 |
| if depth == 0: |
| try: |
| return json.loads(answer[start:i + 1]) |
| except ValueError: |
| break |
| start = answer.find("{", start + 1) |
| return None |
|
|
|
|
| def style_suggestions(answer, text): |
| """[(original, replacement)]: style only, and only where `original` is in the text.""" |
| obj = first_json_object(answer) |
| items = obj.get("suggestions") if isinstance(obj, dict) else None |
| out = [] |
| for item in items if isinstance(items, list) else []: |
| if not isinstance(item, dict): |
| continue |
| o, r = item.get("original"), item.get("replacement") |
| if item.get("type") == "style" and isinstance(o, str) and o and isinstance(r, str) and o != r and o in text: |
| out.append((o, r)) |
| return out |
|
|
