app.py

from fastapi import FastAPI, HTTPException
from fastapi.responses import StreamingResponse
from pydantic import BaseModel
from typing import Optional
from llama_cpp import Llama
import os

# ── Model loading ──────────────────────────────────────────────────────────────
MODEL_REPO = "newtechdevng/i_am_a_lawyer"
MODEL_FILE = "llama-3.2-1b-instruct.Q4_K_M.gguf"
SYSTEM_PROMPT = (
    "You are Ambuj, an expert AI assistant specialised in Indian law. "
    "You provide accurate, well-structured legal information based on Indian statutes, "
    "case law, and legal procedures. Always clarify that your responses are for "
    "informational purposes only and not a substitute for professional legal advice."
)

print("Loading model …")
llm = Llama.from_pretrained(
    repo_id=MODEL_REPO,
    filename=MODEL_FILE,
    n_ctx=512,          # ← was 4096 (killed RAM); 512 is enough for legal Q&A
    n_threads=2,        # ← was os.cpu_count(); free tier has 2 vCPUs, use both safely
    n_batch=64,         # ← smaller prompt batch = less peak RAM
    n_gpu_layers=0,     # ← no GPU on free tier, keep at 0
    verbose=False,
)
print("Model ready ✓")

# ── FastAPI app ────────────────────────────────────────────────────────────────
app = FastAPI(
    title="Indian Legal AI API",
    description="API for the Ambuj Indian Legal Llama model",
    version="1.0.0",
)


# ── Request / Response schemas ─────────────────────────────────────────────────
class Message(BaseModel):
    role: str        # "user" | "assistant" | "system"
    content: str


class ChatRequest(BaseModel):
    messages: list[Message]
    max_tokens: Optional[int] = 256          # ← was 512; lowered default
    temperature: Optional[float] = 0.7
    stream: Optional[bool] = False


class ChatResponse(BaseModel):
    role: str = "assistant"
    content: str


# ── Routes ─────────────────────────────────────────────────────────────────────
@app.get("/")
def root():
    return {
        "name": "Indian Legal AI API",
        "model": MODEL_FILE,
        "endpoints": {
            "POST /chat":   "Send messages, get a response",
            "POST /ask":    "Simple single-question shortcut",
            "GET  /health": "Health check",
            "GET  /docs":   "Swagger UI",
        }
    }


@app.get("/health")
def health():
    return {"status": "ok", "model_loaded": llm is not None}


@app.post("/chat")
def chat(request: ChatRequest):
    # Hard cap max_tokens to prevent OOM on long generations
    safe_tokens = min(request.max_tokens or 256, 256)

    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
    for m in request.messages:
        if m.role not in ("user", "assistant", "system"):
            raise HTTPException(status_code=400, detail=f"Invalid role: {m.role}")
        messages.append({"role": m.role, "content": m.content})

    if request.stream:
        def generate():
            for chunk in llm.create_chat_completion(
                messages=messages,
                max_tokens=safe_tokens,
                temperature=request.temperature,
                stream=True,
            ):
                delta = chunk["choices"][0]["delta"].get("content", "")
                if delta:
                    yield delta

        return StreamingResponse(generate(), media_type="text/plain")

    response = llm.create_chat_completion(
        messages=messages,
        max_tokens=safe_tokens,
        temperature=request.temperature,
        stream=False,
    )
    content = response["choices"][0]["message"]["content"]
    return ChatResponse(content=content)


class AskRequest(BaseModel):
    question: str
    max_tokens: Optional[int] = 256          # ← was 512; lowered default
    temperature: Optional[float] = 0.7


@app.post("/ask")
def ask(request: AskRequest):
    # Hard cap max_tokens to prevent OOM on long generations
    safe_tokens = min(request.max_tokens or 256, 256)

    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user",   "content": request.question},
    ]
    response = llm.create_chat_completion(
        messages=messages,
        max_tokens=safe_tokens,
        temperature=request.temperature,
        stream=False,
    )
    content = response["choices"][0]["message"]["content"]
    return {"question": request.question, "answer": content}