Spaces:
Sleeping
Sleeping
| """Documentary mode, following the 'Ultimate AI Documentary Master Prompt' rules, automated: | |
| research (research.py) -> blueprint + research status + chunked script -> fact check against sources | |
| -> one 4K cinematic photoreal image prompt per chunk, with continuity and shot variety. | |
| The PDF's interactive intake (questions 1-6) maps to UI controls (topic, duration, language, audience, | |
| tone); its NEXT/STOP phases run here one after another. Every function takes `ask(system, user, | |
| temperature, max_tokens) -> dict` so it can be tested without a real model. | |
| """ | |
| from __future__ import annotations | |
| import re | |
| TONES = ["Cinematic & Dramatic", "Serious & Investigative", "Educational & Informative", | |
| "Emotional & Human-Centered", "Suspenseful & Tense", "Balanced Professional Documentary"] | |
| AUDIENCES = ["General Audience", "History Enthusiasts", "Students / Educational", | |
| "YouTube General Audience", "Documentary / Film Audience"] | |
| FUNCTIONS = ["Hook", "Setup", "Development", "Tension", "Turning Point", "Climax", "Aftermath", "Closing"] | |
| SHOTS = ["extreme wide", "wide", "establishing", "medium", "close-up", "extreme close-up", "low angle", | |
| "high angle", "over-the-shoulder", "environmental portrait", "detail", "tracking"] | |
| FOUR_K = "4K cinematic, photorealistic live-action historical documentary" | |
| MAX_WORDS = 22 | |
| GENERIC_PHRASES = ["Nobody could have imagined what happened next", "Everything was about to change", | |
| "A disaster was waiting to happen", "The world would never be the same", | |
| "The situation became increasingly dangerous", "It was a moment in history"] | |
| def chunk_count(seconds: int, words_per_second: float, max_chunks: int) -> int: | |
| """PDF: one 16-21 word chunk per visual unit (1 minute = 6 chunks at a slow English pace). | |
| Scaled by the narrator's real speaking rate so the video fills the chosen length.""" | |
| return max(4, min(max_chunks, round(seconds * words_per_second / 19))) | |
| def words(text: str) -> int: | |
| return len(re.sub(r"[*\"'“”]", "", text).split()) | |
| # ------------------------------------------------------------------------------------ 1. script | |
| def write_script(ask, topic: str, research: dict, language_rules: str, seconds: int, n_chunks: int, | |
| tone: str, audience: str) -> dict: | |
| system = f"""You are a professional historical documentary researcher and scriptwriter. | |
| RESEARCH FIRST, STORY SECOND. You are given SOURCE TEXT. Use ONLY facts that appear in it. | |
| Never invent: dates, times, numbers, names, places, quotes, statistics, dialogue, private thoughts, emotions, | |
| motivations, secret meetings, heroic or villainous behaviour. If the sources are vague or disagree, say so honestly | |
| ("approximately", "records indicate", "historians generally agree", "the exact figure remains disputed"). | |
| Never create false precision. Never present a reconstruction as confirmed fact. | |
| PROCESS: from the sources establish WHEN -> WHERE -> WHO -> WHAT -> HOW -> WHY -> CONSEQUENCE, then write. | |
| NARRATION RULES | |
| - Exact facts drive the story and are embedded naturally: DATE -> LOCATION -> PEOPLE -> SITUATION -> FACTUAL | |
| DEVELOPMENT -> CONSEQUENCE -> NEXT EVENT. Not a list of statistics. | |
| - Prefer specific verified statements over vague ones ("left Southampton on 10 April 1912 with more than 2,000 | |
| people aboard", not "carried many people"). Every number must serve the story. | |
| - Strict chronology. Move the viewer through the timeline naturally ("On ...", "Four days later ...", "By the evening of ..."). | |
| - Let the facts create the drama, where the evidence supports it: NORMALITY -> CONTEXT -> WARNING -> DOCUMENTED | |
| DEVELOPMENT -> ESCALATION -> CONSEQUENCE -> CRISIS -> CLIMAX -> AFTERMATH. Do not force a mystery, twist, | |
| villain or hero the evidence does not support. | |
| - Every chunk must advance the story AND the viewer's knowledge (a fact, a time, a place, a person, a cause, a | |
| consequence). Never spend a chunk only on atmosphere. | |
| - Avoid generic filler such as: {'; '.join(GENERIC_PHRASES)}. | |
| - One professional narrator, continuous story: each chunk follows naturally from the previous one. | |
| - This is a SHORT documentary: essential chronology, major verified facts, key numbers, the turning point, the outcome. | |
| - Chunk 1 is the hook: the most striking VERIFIED fact or question, no clickbait. | |
| - People appear only through documented behaviour. | |
| FORMAT | |
| - Exactly {n_chunks} chunks. Each chunk: one or two complete sentences, maximum {MAX_WORDS} words, preferred 16-21. | |
| - Story functions in order, chosen from: {', '.join(FUNCTIONS)}. | |
| - Tone: {tone}. Audience: {audience}. | |
| - Wrap at most 3 key facts (a date, a number) in *asterisks* for caption highlighting. | |
| {language_rules} | |
| - "title" and "title_options" in Roman-script Hinglish or English, honest, max 60 characters. | |
| Return ONLY JSON: | |
| {{"blueprint": {{"title": "", "core_question": "", "premise": "one line", "period": "", "locations": [], "story_structure": ""}}, | |
| "research_status": {{"verified_facts": ["max 8, each with its figure/date"], "requiring_verification": [], | |
| "reconstructions": [], "uncertainties": []}}, | |
| "chunks": [{{"chunk": 1, "story_function": "Hook", "narration": "", "visual_intent": "what the viewer should see"}}], | |
| "title_options": ["5 titles"], "description": "2-3 lines for YouTube", "hashtags": ["#shorts", "..."]}}""" | |
| user = f"TOPIC: {topic}\nTARGET RUNTIME: about {seconds} seconds\n\nSOURCE TEXT:\n{research['text']}" | |
| doc = ask(system, user, 0.4, 5000) | |
| chunks = [c for c in (doc.get("chunks") or []) if str(c.get("narration", "")).strip()] | |
| if len(chunks) < 3: | |
| raise ValueError(f"documentary script has only {len(chunks)} usable chunks") | |
| for i, c in enumerate(chunks, start=1): | |
| c["chunk"] = i | |
| c["narration"] = str(c["narration"]).strip() | |
| if c.get("story_function") not in FUNCTIONS: | |
| c["story_function"] = FUNCTIONS[min(i - 1, len(FUNCTIONS) - 1)] | |
| doc["chunks"] = chunks[:n_chunks + 2] # tolerate +2, never pad | |
| bp = doc.get("blueprint") or {} | |
| doc["title"] = str(bp.get("title") or doc.get("title") or topic).strip() | |
| return doc | |
| # ------------------------------------------------------------------------------------ 2. fact check | |
| def fact_check(ask, research: dict, doc: dict, language_rules: str) -> tuple[dict, list[dict]]: | |
| system = f"""You are a strict documentary fact-checker. Compare every narration chunk with the SOURCE TEXT. | |
| Flag: any date, time, number, name, place, quote, cause, sequence or claim that the source does not support, | |
| anything stated with more precision than the source gives, invented dialogue/thoughts/emotions, wrong chronology. | |
| Rewrite ONLY the flawed chunks so they are supported by the source (hedge if needed). Keep the same language and | |
| style, maximum {MAX_WORDS} words, keep *asterisks* style. Keep the chunk count and order. | |
| {language_rules} | |
| Return ONLY JSON: {{"issues": [{{"chunk": 1, "problem": "", "fix": ""}}], | |
| "chunks": [{{"chunk": 1, "narration": "all chunks, unchanged ones copied exactly"}}]}}""" | |
| script = "\n".join(f"CHUNK {c['chunk']}: {c['narration']}" for c in doc["chunks"]) | |
| res = ask(system, f"SOURCE TEXT:\n{research['text']}\n\nSCRIPT:\n{script}", 0.1, 4000) | |
| issues = [i for i in (res.get("issues") or []) if isinstance(i, dict)] | |
| fixed = res.get("chunks") or [] | |
| if len(fixed) == len(doc["chunks"]): | |
| for c, f in zip(doc["chunks"], fixed): | |
| new = str(f.get("narration", "")).strip() | |
| if new and new != c["narration"]: | |
| c["narration_before_check"] = c["narration"] | |
| c["narration"] = new | |
| return doc, issues | |
| # ------------------------------------------------------------------------------------ 3. image prompts | |
| def image_prompts(ask, doc: dict, research: dict, aspect: str = "vertical 9:16") -> dict: | |
| system = f"""You are the director of photography of a premium historical documentary. | |
| Write exactly ONE image prompt per chunk. Every prompt MUST begin with: "{FOUR_K}". | |
| Each prompt establishes: exact subject, exact action, historical period, location, characters and clothing, | |
| foreground / midground / background, camera angle and shot size, lens feeling, depth of field, lighting, | |
| atmosphere, color grading, materials. 60-110 words, English. | |
| RULES | |
| - Historical period lock: no modern buildings, vehicles, technology, clothing, weapons, signage or lighting | |
| unless they belong to the scene. Geographic lock: architecture, terrain, vegetation and climate match the place. | |
| - No text, letters, numbers, signs, logos, headlines or typography in the image. | |
| - Global visual tone: every image belongs to the same film - one consistent grade, contrast, lens and | |
| lighting philosophy. No painterly, cartoon, CGI, vintage or oversaturated looks. | |
| - Realism: realistic anatomy, natural skin texture, realistic fabrics, shadows, reflections and scale, natural light. | |
| - Continuity: define recurring characters, locations and props once in "visual_world" and repeat their | |
| appearance word-for-word in every prompt where they appear. | |
| - Cinematography variety: choose a shot from {', '.join(SHOTS)}; never the same shot size twice in a row; | |
| open with an establishing or extreme wide shot. | |
| - Human cost is shown with restraint (aftermath, empty objects, distance, scale) - no gore or graphic injury. | |
| - Composition: {aspect}, main subject readable in the centre third. | |
| Return ONLY JSON: {{"visual_world": {{"period": "", "location": "", "look": "grade, lens, light", | |
| "recurring_characters": [{{"name": "", "appearance": "", "clothing": ""}}], "recurring_props": []}}, | |
| "images": [{{"chunk": 1, "shot": "establishing", "prompt": "{FOUR_K}, ..."}}]}}""" | |
| blueprint = doc.get("blueprint") or {} | |
| plan = "\n".join(f"CHUNK {c['chunk']} [{c['story_function']}]: {c['narration']}\n VISUAL INTENT: " | |
| f"{c.get('visual_intent', '')}" for c in doc["chunks"]) | |
| user = (f"DOCUMENTARY: {doc['title']}\nPERIOD: {blueprint.get('period', '')}\n" | |
| f"LOCATIONS: {', '.join(blueprint.get('locations') or [])}\n\n{plan}\n\n" | |
| f"SOURCE EXCERPT (for period-accurate detail):\n{research['text'][:5000]}") | |
| res = ask(system, user, 0.5, 5000) | |
| images = {int(i.get("chunk", 0)): i for i in (res.get("images") or []) if isinstance(i, dict)} | |
| out = [] | |
| for c in doc["chunks"]: | |
| img = images.get(c["chunk"]) or {} | |
| prompt = str(img.get("prompt") or c.get("visual_intent") or c["narration"]).strip() | |
| if not prompt.lower().startswith("4k"): | |
| prompt = f"{FOUR_K}, {prompt}" | |
| shot = str(img.get("shot", "")).lower() | |
| out.append({"chunk": c["chunk"], "shot": shot if shot in SHOTS else "", "prompt": prompt}) | |
| return {"visual_world": res.get("visual_world") or {}, "images": out} | |
| def camera_motion(shot: str, i: int) -> str: | |
| """PDF motion principles: slow push-in, gentle pan, controlled tracking - no excessive zoom.""" | |
| if shot in ("extreme wide", "establishing", "wide", "tracking"): | |
| return "pan_right" if i % 2 else "pan_left" | |
| if shot == "low angle": | |
| return "pan_up" | |
| if shot == "high angle": | |
| return "pan_down" | |
| if shot in ("close-up", "extreme close-up", "detail", "medium", "environmental portrait", "over-the-shoulder"): | |
| return "zoom_in" | |
| return ["zoom_in", "pan_right", "zoom_out", "pan_left"][i % 4] | |
| def to_board(doc: dict, visuals: dict, clean_tts) -> dict: | |
| scenes = [] | |
| for i, (c, img) in enumerate(zip(doc["chunks"], visuals["images"]), start=1): | |
| scenes.append({"scene_number": i, "narration": c["narration"], "tts_text": clean_tts(c["narration"]), | |
| "story_function": c["story_function"], "visual_prompt": c.get("visual_intent", ""), | |
| "final_prompt": img["prompt"] + ", vertical 9:16 frame, no text", | |
| "shot": img["shot"], "camera_motion": camera_motion(img["shot"], i), | |
| "mood": "mysterious", "sfx": "none"}) | |
| return {"character": (visuals.get("visual_world") or {}).get("recurring_characters") or {}, | |
| "visual_style": (visuals.get("visual_world") or {}).get("look") or FOUR_K, | |
| "visual_world": visuals.get("visual_world") or {}, "music_mood": "mysterious", "scenes": scenes} | |
| def story_view(doc: dict, research: dict, issues: list[dict]) -> dict: | |
| """Shape the documentary like a story dict so the rest of the pipeline and UI stay the same.""" | |
| bp = doc.get("blueprint") or {} | |
| return {"mode": "documentary", "title": doc["title"], "hook": doc["chunks"][0]["narration"], | |
| "story": " ".join(c["narration"] for c in doc["chunks"]), "twist": bp.get("premise", ""), | |
| "ending": doc["chunks"][-1]["narration"], "concept": bp.get("premise", ""), | |
| "blueprint": bp, "research_status": doc.get("research_status") or {}, | |
| "fact_check_issues": issues, "sources": research["sources"], | |
| "title_options": doc.get("title_options") or [doc["title"]], | |
| "description": doc.get("description", ""), "hashtags": doc.get("hashtags") or ["#shorts", "#history"], | |
| "chunks": doc["chunks"]} | |