modal_app.py

import modal
import json
import os
import base64
from pathlib import Path

app = modal.App("memory-keeper")

# ── Volumes ───────────────────────────────────────────────────────────────────
volume = modal.Volume.from_name("memory-keeper-personas", create_if_missing=True)
hf_cache_vol = modal.Volume.from_name("memory-keeper-hf-cache", create_if_missing=True)

# ── MiniCPM Hosted API ────────────────────────────────────────────────────────
MINICPM_LLM_URL    = "http://35.203.155.71:8001/v1"   # MiniCPM4.1-8B
MINICPM_VISION_URL = "http://35.203.155.71:8003/v1"   # MiniCPM-V-4.6
MINICPM_AUTH       = "Bearer sk-minicpm-V8bcD-YTAMxECagaKOnbwTCN69IlN2LhSezGOgq2Ues"
MINICPM_LLM_MODEL  = "MiniCPM4.1-8B"
MINICPM_VIS_MODEL  = "MiniCPM-V-4.6"

# ── Model IDs (Modal-hosted) ──────────────────────────────────────────────────
ASR_MODEL       = "CohereLabs/cohere-transcribe-03-2026"
OCR_MODEL       = "nvidia/NVIDIA-Nemotron-Parse-v1.2"
TTS_MODEL       = "openbmb/VoxCPM2"
MULTILINGUAL_LM_FIRE  = "CohereLabs/tiny-aya-fire"   # South Asian (Bengali, Hindi, Urdu)
MULTILINGUAL_LM_WATER = "CohereLabs/tiny-aya-water"  # Asia Pacific

MINUTES = 60

# ── Images ────────────────────────────────────────────────────────────────────

base_image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install("openai", "requests", "fastapi[standard]", "huggingface_hub")
)

asr_image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(
        "torch", "transformers>=4.50.0", "torchaudio",
        "huggingface_hub", "soundfile", "fastapi[standard]",
    )
)

ocr_image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(
        "torch", "torchvision", "transformers>=4.50.0",
        "Pillow", "huggingface_hub", "fastapi[standard]",
    )
)

tts_image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(
        "torch", "soundfile", "huggingface_hub", "fastapi[standard]",
        "voxcpm",
    )
)

multilingual_image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(
        "torch", "transformers>=4.50.0",
        "huggingface_hub", "fastapi[standard]",
    )
)


# ── Helper: MiniCPM client ────────────────────────────────────────────────────

def get_llm_client():
    from openai import OpenAI
    return OpenAI(
        base_url=MINICPM_LLM_URL,
        api_key=MINICPM_AUTH.replace("Bearer ", ""),
    )

def get_vision_client():
    from openai import OpenAI
    return OpenAI(
        base_url=MINICPM_VISION_URL,
        api_key=MINICPM_AUTH.replace("Bearer ", ""),
    )


# ─────────────────────────────────────────────────────────────────────────────
# 1. ASR — Cohere Transcribe 2B
# ─────────────────────────────────────────────────────────────────────────────

@app.function(gpu="T4", image=asr_image, timeout=5 * MINUTES,
              volumes={"/root/.cache/huggingface": hf_cache_vol})
def transcribe_audio(audio_bytes: bytes, filename: str = "audio.wav") -> str:
    import tempfile, torch, soundfile as sf
    from transformers import pipeline

    pipe = pipeline(
        "automatic-speech-recognition",
        model=ASR_MODEL,
        trust_remote_code=True,
        device="cuda",
        torch_dtype=torch.float16,
    )

    with tempfile.NamedTemporaryFile(suffix=Path(filename).suffix, delete=False) as f:
        f.write(audio_bytes)
        tmp_path = f.name

    result = pipe(tmp_path, return_timestamps=False)
    os.unlink(tmp_path)
    return result["text"] if isinstance(result, dict) else str(result)


# ─────────────────────────────────────────────────────────────────────────────
# 2. OCR — Nemotron Parse v1.2
# ─────────────────────────────────────────────────────────────────────────────

@app.function(gpu="T4", image=ocr_image, timeout=5 * MINUTES,
              volumes={"/root/.cache/huggingface": hf_cache_vol})
def ocr_document(image_bytes: bytes) -> str:
    import tempfile, torch
    from PIL import Image
    from transformers import AutoModelForImageTextToText, AutoProcessor

    processor = AutoProcessor.from_pretrained(OCR_MODEL, trust_remote_code=True)
    model = AutoModelForImageTextToText.from_pretrained(
        OCR_MODEL, trust_remote_code=True,
        torch_dtype=torch.float16, device_map="auto",
    )
    model.eval()

    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as f:
        f.write(image_bytes)
        tmp_path = f.name

    image = Image.open(tmp_path).convert("RGB")
    os.unlink(tmp_path)

    messages = [{"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Extract all text from this document. Preserve structure and formatting."},
    ]}]
    inputs = processor.apply_chat_template(
        messages, tokenize=True, add_generation_prompt=True,
        return_dict=True, return_tensors="pt",
    ).to(model.device)

    generated_ids = model.generate(**inputs, max_new_tokens=2048)
    trimmed = [out[len(inp):] for inp, out in zip(inputs.input_ids, generated_ids)]
    return processor.batch_decode(trimmed, skip_special_tokens=True)[0]


# ─────────────────────────────────────────────────────────────────────────────
# 3. TTS — VoxCPM2
# ─────────────────────────────────────────────────────────────────────────────

@app.function(gpu="T4", image=tts_image, timeout=5 * MINUTES,
              volumes={"/root/.cache/huggingface": hf_cache_vol})
def text_to_speech(text: str, voice_description: str = None) -> bytes:
    import soundfile as sf, io
    from voxcpm import VoxCPM

    model = VoxCPM.from_pretrained(TTS_MODEL, load_denoiser=False)
    if voice_description:
        text = f"({voice_description}){text}"

    wav = model.generate(text=text, cfg_value=2.0, inference_timesteps=10)
    buf = io.BytesIO()
    sf.write(buf, wav, model.tts_model.sample_rate, format="WAV")
    return buf.getvalue()


# ─────────────────────────────────────────────────────────────────────────────
# 4. Multilingual — Tiny Aya Fire (Bengali/South Asian fallback)
# ─────────────────────────────────────────────────────────────────────────────

@app.function(gpu="T4", image=multilingual_image, timeout=5 * MINUTES,
              volumes={"/root/.cache/huggingface": hf_cache_vol})
def chat_multilingual(system_prompt: str, history: list, user_message: str, model_id: str = None) -> str:
    import torch
    from transformers import AutoTokenizer, AutoModelForCausalLM

    if model_id is None:
        model_id = MULTILINGUAL_LM_FIRE
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    model = AutoModelForCausalLM.from_pretrained(
        model_id, torch_dtype=torch.float16, device_map="auto",
    )
    model.eval()

    messages = [{"role": "system", "content": system_prompt}]
    for msg in history[-8:]:
        messages.append({"role": msg["role"], "content": msg["content"]})
    messages.append({"role": "user", "content": user_message})

    input_ids = tokenizer.apply_chat_template(
        messages, tokenize=True, add_generation_prompt=True, return_tensors="pt",
    ).to(model.device)

    gen_tokens = model.generate(
        input_ids, max_new_tokens=512, do_sample=True,
        temperature=0.8, top_p=0.95,
    )
    output = gen_tokens[0][input_ids.shape[-1]:]
    return tokenizer.decode(output, skip_special_tokens=True)


# ─────────────────────────────────────────────────────────────────────────────
# 5. Photo Description — MiniCPM-V 4.6 (hosted API)
# ─────────────────────────────────────────────────────────────────────────────

@app.function(image=base_image, timeout=5 * MINUTES)
def describe_photo(image_bytes: bytes, prompt: str = None) -> str:
    import base64
    client = get_vision_client()

    img_b64 = base64.b64encode(image_bytes).decode()
    if prompt is None:
        prompt = (
            "Describe this person in detail: their appearance, expression, what they are doing, "
            "the setting, and any emotional tone you sense. "
            "This will help reconstruct their memory and personality."
        )

    response = client.chat.completions.create(
        model=MINICPM_VIS_MODEL,
        messages=[{"role": "user", "content": [
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img_b64}"}},
            {"type": "text", "text": prompt},
        ]}],
        max_tokens=512,
        temperature=0.3,
    )
    return response.choices[0].message.content.strip()


# ─────────────────────────────────────────────────────────────────────────────
# 6. Persona Builder — MiniCPM4.1-8B (hosted API)
# ─────────────────────────────────────────────────────────────────────────────

@app.function(image=base_image, timeout=20 * MINUTES, volumes={"/personas": volume})
def build_persona(
    persona_id: str, name: str, relationship: str,
    texts: list, photo_captions: list, voice_transcripts: list,
) -> dict:
    client = get_llm_client()

    all_content = []
    if texts:
        all_content.append("=== LETTERS / DIARY ENTRIES ===\n" + "\n---\n".join(texts))
    if photo_captions:
        all_content.append("=== PHOTO DESCRIPTIONS ===\n" + "\n---\n".join(photo_captions))
    if voice_transcripts:
        all_content.append("=== VOICE TRANSCRIPTS ===\n" + "\n---\n".join(voice_transcripts))

    combined = "\n\n".join(all_content)

    prompt = f"""You are helping preserve the memory of {name} ({relationship}).

Below is everything we have from them:

{combined}

Extract a rich persona profile. Return ONLY valid JSON, no markdown, no backticks:
{{
  "name": "{name}",
  "relationship": "{relationship}",
  "personality_traits": ["5-8 traits"],
  "speech_style": "how they speak, tone, vocabulary, sentence length",
  "common_phrases": ["phrases or expressions they often used"],
  "key_memories": ["10-15 specific memories or life events"],
  "values": ["what they cared about most"],
  "language": "primary language (English/Bengali/Hindi/mixed)",
  "emotional_tone": "overall emotional quality",
  "topics_they_loved": ["subjects they talked about often"],
  "advice_they_gave": ["wisdom or advice they shared"],
  "voice_description": "describe their voice: age, gender, tone, accent e.g. elderly Bengali man, warm gentle voice",
  "system_prompt": "A 300-word system prompt starting with: You are {name}..."
}}"""

    response = client.chat.completions.create(
        model=MINICPM_LLM_MODEL,
        messages=[{"role": "user", "content": prompt}],
        temperature=0.3,
        max_tokens=2048,
    )

    import re
    raw = response.choices[0].message.content.strip()
    print(f"RAW RESPONSE (first 500 chars): {raw[:500]}")

    # Strip <think>...</think> tags
    raw = re.sub(r'<think>.*?</think>', '', raw, flags=re.DOTALL).strip()
    print(f"AFTER THINK STRIP (first 500 chars): {raw[:500]}")

    # Strip markdown code blocks
    if "```" in raw:
        parts = raw.split("```")
        for part in parts:
            if part.startswith("json"):
                part = part[4:]
            part = part.strip()
            if part.startswith("{"):
                raw = part
                break

    # Extract JSON object
    json_match = re.search(r'[{].*[}]', raw, re.DOTALL)
    if json_match:
        raw = json_match.group(0)

    print(f"FINAL RAW FOR PARSING (first 300 chars): {raw[:300]}")
    persona = json.loads(raw)
    os.makedirs("/personas", exist_ok=True)
    with open(f"/personas/{persona_id}.json", "w", encoding="utf-8") as f:
        json.dump(persona, f, ensure_ascii=False, indent=2)
    volume.commit()
    return persona


# ─────────────────────────────────────────────────────────────────────────────
# 7. Chat — MiniCPM4.1-8B (hosted API)
# ─────────────────────────────────────────────────────────────────────────────

BENGALI_CHARS = set("অআইঈউঊঋএঐওঔকখগঘঙচছজঝঞটঠডঢণতথদধনপফবভমযরলশষসহড়ঢ়য়ংঃ")
# Asia Pacific: Chinese, Japanese, Korean, Thai, Vietnamese tones etc.
ASIA_PACIFIC_RANGES = [(0x4E00, 0x9FFF), (0x3040, 0x30FF), (0xAC00, 0xD7AF), (0x0E00, 0x0E7F)]

def is_asia_pacific(text):
    for ch in text:
        cp = ord(ch)
        if any(lo <= cp <= hi for lo, hi in ASIA_PACIFIC_RANGES):
            return True
    return False

@app.function(image=base_image, timeout=5 * MINUTES, volumes={"/personas": volume})
def chat_with_persona(
    persona_id: str, history: list, user_message: str, language: str = "auto",
) -> dict:
    persona_path = f"/personas/{persona_id}.json"
    if not os.path.exists(persona_path):
        raise FileNotFoundError(f"Persona '{persona_id}' not found.")

    with open(persona_path, "r", encoding="utf-8") as f:
        persona = json.load(f)

    system_prompt = persona.get("system_prompt", f"You are {persona['name']}.")
    voice_desc = persona.get("voice_description", "warm elderly voice")

    full_system = f"""{system_prompt}

IMPORTANT:
- You ARE {persona['name']}. Never break character.
- Use their speech style, phrases, and memories naturally.
- Be warm and personal, not like an AI.
- Detect the user's language and respond in the same language.
- Keep responses 2-4 sentences unless sharing a story."""

    import re

    is_bengali = any(c in BENGALI_CHARS for c in user_message)
    is_ap = is_asia_pacific(user_message)

    # All languages through MiniCPM4.1-8B (handles Bengali/Hindi/Chinese natively)
    # Tiny Aya Fire/Water kept for dedicated multilingual endpoint (future use)
    client = get_llm_client()

    # Add language instruction to system prompt
    if is_bengali or language == "Bengali":
        full_system = full_system + "\n\nIMPORTANT: The user is writing in Bengali. You MUST respond in Bengali only."
    elif is_ap or language in ("Chinese", "Japanese", "Korean", "Thai"):
        full_system = full_system + "\n\nIMPORTANT: Respond in " + str(language) + " only."


    messages = [{"role": "system", "content": full_system}]
    for msg in history[-10:]:
        messages.append({"role": msg["role"], "content": msg["content"]})
    messages.append({"role": "user", "content": user_message})
    response = client.chat.completions.create(
        model=MINICPM_LLM_MODEL,
        messages=messages,
        temperature=0.8,
        max_tokens=1024,
        stop=None,
    )
    response_text = response.choices[0].message.content.strip()
    response_text = re.sub(r'<think>.*?</think>', '', response_text, flags=re.DOTALL).strip()
    # Remove any truncated incomplete sentence at end
    if response_text and not response_text[-1] in '.!?।':
        last_punct = max(
            response_text.rfind('.'),
            response_text.rfind('!'),
            response_text.rfind('?'),
            response_text.rfind('।'),
        )
        if last_punct > len(response_text) // 2:
            response_text = response_text[:last_punct+1]

    return {"text": response_text, "voice_description": voice_desc}


# ─────────────────────────────────────────────────────────────────────────────
# 8. Web Endpoints
# ─────────────────────────────────────────────────────────────────────────────

from fastapi import FastAPI, Request
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import Response


def make_app():
    web_app = FastAPI()
    web_app.add_middleware(
        CORSMiddleware, allow_origins=["*"],
        allow_methods=["*"], allow_headers=["*"],
    )
    return web_app


# Health check — tests MiniCPM API reachability
health_web = make_app()
@app.function(image=base_image, timeout=30)
@modal.asgi_app(label="health")
def health_endpoint():
    @health_web.get("/")
    async def handler():
        try:
            client = get_llm_client()
            r = client.chat.completions.create(
                model=MINICPM_LLM_MODEL,
                messages=[{"role": "user", "content": "Hi"}],
                max_tokens=5,
            )
            return {"status": "ok", "minicpm": "reachable", "reply": r.choices[0].message.content}
        except Exception as e:
            return {"status": "error", "minicpm": "unreachable", "error": str(e)}
    return health_web


# Build persona
build_web = make_app()
@app.function(image=base_image, volumes={"/personas": volume}, timeout=20 * MINUTES)
@modal.asgi_app(label="build-persona")
def build_persona_endpoint():
    @build_web.post("/")
    async def handler(request: Request):
        try:
            data = await request.json()
            persona = build_persona.remote(
                persona_id=data["persona_id"], name=data["name"],
                relationship=data["relationship"], texts=data.get("texts", []),
                photo_captions=data.get("photo_captions", []),
                voice_transcripts=data.get("voice_transcripts", []),
            )
            return {"success": True, "persona": persona}
        except Exception as e:
            return {"success": False, "error": str(e)}
    return build_web


# Chat
chat_web = make_app()
@app.function(image=base_image, volumes={"/personas": volume}, timeout=5 * MINUTES)
@modal.asgi_app(label="chat")
def chat_endpoint():
    @chat_web.post("/")
    async def handler(request: Request):
        data = await request.json()
        result = chat_with_persona.remote(
            persona_id=data["persona_id"], history=data.get("history", []),
            user_message=data["message"], language=data.get("language", "auto"),
        )
        return result
    return chat_web


# Transcribe
transcribe_web = make_app()
@app.function(image=asr_image, volumes={"/root/.cache/huggingface": hf_cache_vol}, timeout=5 * MINUTES)
@modal.asgi_app(label="transcribe")
def transcribe_endpoint():
    @transcribe_web.post("/")
    async def handler(request: Request):
        data = await request.json()
        audio_bytes = base64.b64decode(data["audio_b64"])
        transcript = transcribe_audio.remote(audio_bytes, data.get("filename", "audio.wav"))
        return {"transcript": transcript}
    return transcribe_web


# Vision
vision_web = make_app()
@app.function(image=base_image, timeout=5 * MINUTES)
@modal.asgi_app(label="describe-photo")
def vision_endpoint():
    @vision_web.post("/")
    async def handler(request: Request):
        data = await request.json()
        image_bytes = base64.b64decode(data["image_b64"])
        description = describe_photo.remote(image_bytes, data.get("prompt"))
        return {"description": description}
    return vision_web


# OCR
ocr_web = make_app()
@app.function(image=ocr_image, volumes={"/root/.cache/huggingface": hf_cache_vol}, timeout=5 * MINUTES)
@modal.asgi_app(label="ocr-document")
def ocr_endpoint():
    @ocr_web.post("/")
    async def handler(request: Request):
        data = await request.json()
        image_bytes = base64.b64decode(data["image_b64"])
        text = ocr_document.remote(image_bytes)
        return {"text": text}
    return ocr_web


# TTS
tts_web = make_app()
@app.function(image=tts_image, volumes={"/root/.cache/huggingface": hf_cache_vol}, timeout=5 * MINUTES)
@modal.asgi_app(label="text-to-speech")
def tts_endpoint():
    @tts_web.post("/")
    async def handler(request: Request):
        data = await request.json()
        wav_bytes = text_to_speech.remote(data["text"], data.get("voice_description"))
        return Response(content=wav_bytes, media_type="audio/wav")
    return tts_web


# List personas
list_web = make_app()
@app.function(image=base_image, volumes={"/personas": volume}, timeout=MINUTES)
@modal.asgi_app(label="list-personas")
def list_personas_endpoint():
    @list_web.get("/")
    async def handler():
        personas = []
        if os.path.exists("/personas"):
            for f in Path("/personas").glob("*.json"):
                with open(f) as fp:
                    data = json.load(fp)
                    personas.append({
                        "id": f.stem,
                        "name": data.get("name"),
                        "relationship": data.get("relationship"),
                    })
        return {"personas": personas}
    return list_web