File size: 8,121 Bytes
a654c7f
 
63bc04d
a654c7f
 
 
 
 
 
63bc04d
a654c7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63bc04d
a654c7f
 
 
63bc04d
a654c7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
63bc04d
 
 
 
 
 
 
 
ef36da0
 
 
 
a654c7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
import os
import sys
import base64
import urllib.parse
from datetime import datetime
from typing import Any
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from fastapi.staticfiles import StaticFiles
from fastapi.responses import FileResponse, Response
from pydantic import BaseModel, Field
import uvicorn

# --- Core dependencies ---
try:
    from llama_cpp import Llama
    print("βœ… llama-cpp-python")
except ImportError:
    print("❌ Run: pip install llama-cpp-python")
    sys.exit(1)

from rag_utils import (
    ABSTAIN_MESSAGE,
    build_general_system_prompt,
    build_hybrid_system_prompt,
    build_system_prompt,
    compose_krce_response,
    finalize_general_response,
    finalize_krce_response,
    load_rag_index,
    search_krce,
)

# --- Config ---
# Model settings
REPO_ID = "Krishkanth/krish-mind-mobile"
MODEL_FILENAME = "krish-mind-mobile.gguf"
BASE_DIR = os.path.dirname(__file__)
STATIC_DIR = os.path.join(BASE_DIR, "static")
LOGO_B64_FILE = os.path.join(STATIC_DIR, "logo_png_base64.txt")
default_clean_data = os.path.join(BASE_DIR, "data", "krce_college_data_clean.jsonl")
default_legacy_data = os.path.join(BASE_DIR, "data", "krce_college_data.jsonl")
DATA_FILE = default_clean_data if os.path.exists(default_clean_data) else default_legacy_data
_logo_png_cache: bytes | None = None

# --- Load GGUF Model ---
print(f"\n⏳ Downloading/Loading model from {REPO_ID}...")
try:
    from huggingface_hub import hf_hub_download
    
    # Download model (cached)
    model_path = hf_hub_download(
        repo_id=REPO_ID,
        filename=MODEL_FILENAME,
        local_dir="model", # Download to local folder
        local_dir_use_symlinks=False
    )
    print(f"βœ… Model downloaded to: {model_path}")

    model = Llama(
        model_path=model_path,
        n_ctx=4096,
        n_gpu_layers=0, # CPU only for free tier
        verbose=False
    )
    print("βœ… Model loaded!")

except Exception as e:
    print(f"❌ Model error: {e}")
    model = None

# --- RAG SETUP ---
print("πŸ“š Indexing Knowledge Base...")
rag_index = load_rag_index(DATA_FILE)
if rag_index.model is not None and rag_index.records:
    print(f"βœ… Indexed {len(rag_index.records)} KRCE facts.")
else:
    print("⚠️ Data file not found or embedding model unavailable. RAG disabled.")

# --- FastAPI ---
app = FastAPI()
app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])

# Serve Static Files
app.mount("/static", StaticFiles(directory=STATIC_DIR), name="static")

class ChatRequest(BaseModel):
    message: str
    max_tokens: int = 1024
    temperature: float = 0.1
    krce_mode: bool = False
    history: list[dict[str, Any]] = Field(default_factory=list)

@app.get("/")
async def root():
    # Serve index.html at root
    return FileResponse(os.path.join(STATIC_DIR, "index.html"))

@app.get("/logo.png")
async def logo():
    global _logo_png_cache
    if _logo_png_cache is None:
        if os.path.exists(LOGO_B64_FILE):
            with open(LOGO_B64_FILE, "r", encoding="ascii") as handle:
                _logo_png_cache = base64.b64decode(handle.read().strip())
        else:
            return FileResponse(os.path.join(STATIC_DIR, "logo.svg"), media_type="image/svg+xml")
    return Response(content=_logo_png_cache, media_type="image/png")

@app.get("/logo.svg")
async def logo_svg():
    return FileResponse(os.path.join(STATIC_DIR, "logo.svg"), media_type="image/svg+xml")

@app.post("/chat")
async def chat(request: ChatRequest):
    if not model:
        return {"response": "Error: Model not loaded. Please check server logs."}
        
    user_input = request.message
    
    # Image Generation Hook
    if any(t in user_input.lower() for t in ["generate image", "create image", "draw", "imagine"]):
        prompt = user_input.replace("generate image", "").strip()
        url = f"https://image.pollinations.ai/prompt/{urllib.parse.quote(prompt)}"
        return {"response": f"Here's your image of **{prompt}**:\n\n![{prompt}]({url})"}

    # Frontend controls route explicitly:
    # - KRCE mode ON: strict grounded KRCE answers only
    # - KRCE mode OFF: normal model chat without RAG retrieval
    route = "krce" if bool(request.krce_mode) else "general"
    rag_result = {
        "context": "",
        "hits": [],
        "should_abstain": False,
        "confidence": 0.0,
    }

    if route in {"krce", "hybrid"}:
        rag_result = search_krce(user_input, rag_index)
        if rag_result["context"]:
            print(f"\n[πŸ“¦ RAG CONTEXT FOUND]\n{rag_result['context']}\n")

    if route == "krce" and rag_result["should_abstain"]:
        return {"response": ABSTAIN_MESSAGE}

    if route == "krce" and rag_result.get("hits"):
        response_text = compose_krce_response(user_input, rag_result)
        return {"response": finalize_krce_response(user_input, response_text, rag_result)}

    now = datetime.now().strftime("%A, %B %d, %Y")
    if route == "hybrid":
        sys_prompt = build_hybrid_system_prompt(now, rag_result)
    elif route == "general":
        sys_prompt = build_general_system_prompt(now)
    else:
        sys_prompt = build_system_prompt(now, user_input, rag_result)

    prompt_text = user_input
    if route == "general" and request.history:
        compact_turns: list[str] = []
        for turn in request.history[-8:]:
            role = str(turn.get("role", "")).strip().lower()
            content = str(turn.get("content", "")).strip()
            if role not in {"user", "assistant"} or not content:
                continue
            if len(content) > 1200:
                content = content[:1200].rstrip() + " ..."
            speaker = "User" if role == "user" else "Assistant"
            compact_turns.append(f"{speaker}: {content}")
        if compact_turns:
            prompt_text = (
                "Conversation context (most recent turns):\n"
                + "\n".join(compact_turns)
                + "\n\nUser: "
                + user_input
                + "\nAssistant:"
            )

    full_prompt = f"<|system|>\n{sys_prompt}<|end|>\n<|user|>\n{prompt_text}<|end|>\n<|assistant|>\n"

    # Enforce strict stop tokens to prevent the model from hallucinating user prompts or looping
    stop_tokens = ["<|end|>", "<|endoftext|>", "<|user|>", "<|system|>"]
    
    try:
        max_allowed = 420 if route == "krce" else 1200
        effective_tokens = max(64, min(int(request.max_tokens), max_allowed))
        effective_temp = min(request.temperature, 0.1) if route == "krce" else min(max(request.temperature, 0.2), 0.6)

        output = model(
            full_prompt, 
            max_tokens=effective_tokens, 
            temperature=effective_temp,
            repeat_penalty=1.15, # Prevents text repeating/gibberish loops
            stop=stop_tokens, 
            echo=False
        )
        response_text = output["choices"][0]["text"].strip()

        finish_reason = str(output["choices"][0].get("finish_reason", "")).lower()
        if route == "general" and finish_reason == "length" and response_text:
            continue_prompt = (
                f"{full_prompt}{response_text}\n"
                "Continue from where it stopped. Do not repeat previous lines. "
                "Finish the answer clearly."
            )
            cont = model(
                continue_prompt,
                max_tokens=min(400, max_allowed),
                temperature=max(0.15, min(effective_temp, 0.4)),
                repeat_penalty=1.12,
                stop=stop_tokens,
                echo=False,
            )
            extra = cont["choices"][0]["text"].strip()
            if extra:
                response_text = (response_text + "\n" + extra).strip()

        if route == "krce":
            return {"response": finalize_krce_response(user_input, response_text, rag_result)}
        return {"response": finalize_general_response(user_input, response_text)}
    except Exception as e:
        return {"response": f"Error: {e}"}

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)