File size: 5,583 Bytes
6835659 | 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 | from __future__ import annotations
import logging
from dataclasses import dataclass
from typing import Optional
import json
import re
import urllib.request
from src.exceptions import GenerationError
logger = logging.getLogger(__name__)
@dataclass(frozen=True)
class TextGenResult:
text: str
def _sanitize_text(s: str) -> str:
"""Remove common failure patterns (echoing rules, bullets, repetitions)."""
s = s.strip()
# Remove markdown/bullets
s = re.sub(r"^\s*[-*•]\s+", "", s, flags=re.MULTILINE)
# Remove obvious meta/instruction echoes
bad_patterns = [
r"(?i)\blength\s*:\s*\d+\s*[-–]\s*\d+\s*sentences\b.*",
r"(?i)\brules\s*:\b.*",
r"(?i)\bno bullet points\b.*",
r"(?i)\bno repetition\b.*",
r"(?i)\bno meta commentary\b.*",
r"(?i)\bdescribe only\b.*",
]
for pat in bad_patterns:
s = re.sub(pat, "", s).strip()
# Collapse whitespace
s = re.sub(r"\n{3,}", "\n\n", s)
s = re.sub(r"[ \t]{2,}", " ", s)
# If the model repeated the same line many times, de-dup consecutive duplicates
lines = [ln.strip() for ln in s.splitlines() if ln.strip()]
deduped = []
for ln in lines:
if not deduped or deduped[-1] != ln:
deduped.append(ln)
s = "\n".join(deduped).strip()
return s
def _ollama_generate(
prompt: str,
model: str = "qwen2:7b",
temperature: float = 0.7,
top_p: float = 0.9,
num_predict: int = 180,
host: str = "http://localhost:11434",
) -> str:
"""
Calls Ollama local server: POST /api/generate
"""
url = f"{host.rstrip('/')}/api/generate"
payload = {
"model": model,
"prompt": prompt,
"stream": False,
"options": {
"temperature": temperature,
"top_p": top_p,
"num_predict": num_predict,
},
}
req = urllib.request.Request(
url,
data=json.dumps(payload).encode("utf-8"),
headers={"Content-Type": "application/json"},
method="POST",
)
try:
with urllib.request.urlopen(req, timeout=600) as resp:
data = json.loads(resp.read().decode("utf-8"))
text = data.get("response", "").strip()
logger.debug("Ollama generated %d chars", len(text))
return text
except Exception as e:
logger.error("Ollama call failed on %s: %s", host, e)
raise GenerationError(
f"Ollama call failed. Is Ollama running on {host}? Error: {e}",
modality="text", backend=f"ollama/{model}",
) from e
class TextGenerator:
"""
Option A (recommended): Ollama text generator (instruction-following).
Falls back to HF pipeline if use_ollama=False.
"""
def __init__(
self,
use_ollama: bool = True,
ollama_model: str = "qwen2:7b",
ollama_host: str = "http://localhost:11434",
max_new_tokens: int = 160,
hf_model_name: str = "gpt2",
):
self.use_ollama = use_ollama
self.ollama_model = ollama_model
self.ollama_host = ollama_host
self.max_new_tokens = max_new_tokens
self.hf_model_name = hf_model_name
self._hf_pipe = None
if not self.use_ollama:
from transformers import pipeline
self._hf_pipe = pipeline("text-generation", model=self.hf_model_name)
def generate(self, prompt: str, deterministic: bool = True) -> TextGenResult:
# This is the IMPORTANT part: we wrap your plan_text with strict generation rules.
wrapped_prompt = """You are a concise descriptive writer.
Write a literal description of the same scene. Follow these rules:
- Write 3 to 5 natural sentences.
- No bullet points, no numbered lists.
- No repetition.
- No meta commentary (do not mention rules, prompts, or constraints).
- Focus on concrete visual details AND the likely audio ambience.
SCENE PLAN:
"""
wrapped_prompt = f"{wrapped_prompt}{prompt}\n\nNow write the description:\n"
if self.use_ollama:
raw = _ollama_generate(
prompt=wrapped_prompt,
model=self.ollama_model,
host=self.ollama_host,
temperature=0.0 if deterministic else 0.7,
top_p=1.0 if deterministic else 0.9,
num_predict=max(self.max_new_tokens, 120),
)
clean = _sanitize_text(raw)
# Last safety: if it comes out empty, return raw (better than nothing)
return TextGenResult(text=clean if clean else raw)
# HF fallback
outputs = self._hf_pipe(
wrapped_prompt,
max_new_tokens=self.max_new_tokens,
do_sample=not deterministic,
temperature=0.0 if deterministic else 0.9,
top_p=1.0 if deterministic else 0.95,
num_return_sequences=1,
)
text = outputs[0]["generated_text"]
text = _sanitize_text(text)
return TextGenResult(text=text)
def generate_text(
prompt: str,
use_ollama: bool = True,
deterministic: bool = True,
ollama_model: str = "qwen2:7b",
ollama_host: str = "http://localhost:11434",
max_new_tokens: int = 160,
hf_model_name: str = "gpt2",
) -> str:
generator = TextGenerator(
use_ollama=use_ollama,
ollama_model=ollama_model,
ollama_host=ollama_host,
max_new_tokens=max_new_tokens,
hf_model_name=hf_model_name,
)
return generator.generate(prompt, deterministic=deterministic).text
|