drjones/auto-publisher
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: bump llm_chat timeout 60s→600s for qwen3.8 long-form (was timing out on article writing)

Browse Source
This commit is contained in:
drjones
2026-08-15 09:28:07 -07:00
parent 8ce68fa779
commit ec35333926
1 changed files with 20 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
core/orchestrator.py
View File

@@ -16,7 +16,8 @@ import requests
BASE_DIR = Path(__file__).resolve().parent.parent
DB_PATH = BASE_DIR / "core" / "publisher.db"
OLLAMA_MACBOOK = "http://localhost:11434"
OLLAMA_GAMINGPC = "http://10.30.20.186:11434" # RTX 3070, ornith:latest
OLLAMA_GAMINGPC = "http://10.30.20.186:11434" # RTX 3070, ornith:latest (fallback)
OLLAMA_SHADOW = "http://10.30.20.128:11434" # RTX 4080 SUPER, qwen3.8:latest (primary)
# Load API keys from Hermes env if not already set
_hermes_env = Path.home() / ".hermes" / ".env"
@@ -176,7 +177,7 @@ def _call_deepseek(prompt: str, model: str = "deepseek-chat", system: str = "",
raise RuntimeError(f"DeepSeek API error {r.status_code}: {r.text[:200]}")
def llm_chat(prompt: str, model: str = "qwen3.5:4b-mlx", host: str = OLLAMA_MACBOOK,
def llm_chat(prompt: str, model: str = "qwen3.8:latest", host: str = OLLAMA_SHADOW,
system: str = "", temperature: float = 0.7, max_tokens: int = 4096,
retries: int = 3) -> str:
"""Call LLM with DeepSeek cloud → Ollama fallback, with retries."""
@@ -188,18 +189,19 @@ def llm_chat(prompt: str, model: str = "qwen3.5:4b-mlx", host: str = OLLAMA_MACB
log.warning(f"DeepSeek failed, trying local Ollama: {e}")
payload = {
"model": model, "messages": [], "stream": False,
"options": {"temperature": temperature, "num_predict": max_tokens}
"think": False,
"options": {"temperature": temperature, "num_predict": max_tokens, "num_ctx": 8192}
}
if system:
payload["messages"].append({"role": "system", "content": system})
payload["messages"].append({"role": "user", "content": prompt})
# Try Ollama hosts first
hosts = list(dict.fromkeys([host, OLLAMA_MACBOOK, OLLAMA_GAMINGPC]))
hosts = list(dict.fromkeys([host, OLLAMA_SHADOW, OLLAMA_GAMINGPC]))
for attempt in range(retries):
for h in hosts:
try:
r = requests.post(f"{h}/api/chat", json=payload, timeout=60 * (attempt + 1),
r = requests.post(f"{h}/api/chat", json=payload, timeout=600,
proxies={"http": None, "https": None})
if r.status_code == 200:
result = r.json()
@@ -235,7 +237,7 @@ def llm_chat(prompt: str, model: str = "qwen3.5:4b-mlx", host: str = OLLAMA_MACB
raise RuntimeError(f"All LLM hosts failed for model {model}")
def llm_json(prompt: str, model: str = "qwen3.5:4b-mlx", host: str = OLLAMA_MACBOOK,
def llm_json(prompt: str, model: str = "qwen3.8:latest", host: str = OLLAMA_SHADOW,
system: str = "You are a JSON-only API. Always respond with valid JSON. No markdown, no explanation.",
temperature: float = 0.3) -> dict:
"""Call LLM and parse JSON response."""
@@ -252,11 +254,11 @@ def dual_llm_research(prompt: str, system: str = "") -> tuple[str, dict]:
import concurrent.futures
def call_ornith():
return llm_chat(prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
return llm_chat(prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW,
system=system, temperature=0.3, max_tokens=4096)
def call_qwen():
return llm_chat(prompt, model="qwen3.5:4b-mlx", host=OLLAMA_MACBOOK,
return llm_chat(prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW,
system=system, temperature=0.3, max_tokens=2048)
with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
@@ -452,7 +454,7 @@ Cover these verticals: AI/ML, general tech, science, cryptocurrency, Linux, gami
Respond with a JSON array of strings, each a compelling article title."""
try:
result = llm_json(prompt, model="minicpm-v4.5:8b", host=OLLAMA_GAMINGPC, temperature=0.8)
result = llm_json(prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW, temperature=0.8)
if isinstance(result, list):
return result
return list(result.values())[0] if result else []
@@ -641,11 +643,11 @@ Extract and return as JSON:
Be accurate. Cite real sources. No hallucinations. Respond with ONLY valid JSON."""
try:
result = llm_json(research_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
result = llm_json(research_prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW,
system="You are an expert research analyst. You produce accurate, well-cited research. Never fabricate information.")
except Exception as e:
log.error(f"Research LLM failed: {e}. Falling back to MacBook.")
result = llm_json(research_prompt, model="minicpm-v4.5:8b", host=OLLAMA_GAMINGPC,
result = llm_json(research_prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW,
system="You are an expert research analyst. Be accurate and honest.")
# Store knowledge package
@@ -831,8 +833,8 @@ Dark background matching the site's aesthetic. Abstract but relevant to the topi
verify = llm_chat(
f"""Examine this image and verify it's appropriate for an article titled "{title}" on a {vertical} website.
Is the image relevant, coherent, and free of inappropriate content? Respond ONLY with "PASS" or "FAIL: <reason>".""",
model="minicpm-v4.6:1b",
host=OLLAMA_MACBOOK,
model="qwen3.8:latest",
host=OLLAMA_SHADOW,
system="You are an image quality reviewer. Be strict but fair.",
temperature=0.1,
max_tokens=50,
@@ -875,7 +877,7 @@ Generate an outline appropriate for this format.
Respond with JSON:
{{"sections": [{{"heading": "...", "subsections": ["..."]}}, ...], "faq_questions": ["..."], "cta": "..."}}"""
outline = llm_json(outline_prompt, model="minicpm-v4.5:8b", host=OLLAMA_GAMINGPC, temperature=0.5)
outline = llm_json(outline_prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW, temperature=0.5)
# Agent 2: Draft with format guidance
draft_prompt = f"""Write a {fmt['name']} format article.
@@ -902,7 +904,7 @@ Requirements:
Respond with the FULL Markdown article. No JSON wrapper."""
draft = llm_chat(draft_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
draft = llm_chat(draft_prompt, model="qwen3.8:latest", host=OLLAMA_SHADOW,
system="You are an expert writer. Write clear, accurate, engaging content. No AI clichés. No fluff.",
temperature=0.75, max_tokens=8192)
@@ -914,14 +916,14 @@ ARTICLE:
{draft}
Return the edited article in full Markdown. No JSON wrapper.""",
model="minicpm-v4.5:8b", host=OLLAMA_GAMINGPC, temperature=0.3, max_tokens=8192)
model="qwen3.8:latest", host=OLLAMA_SHADOW, temperature=0.3, max_tokens=8192)
# Agent 4: SEO
seo = llm_json(f"""Optimize this article for SEO.
TITLE: {topic_title}
FIRST 500 CHARS: {edited[:500]}
Respond with JSON: {{"seo_title": "...", "seo_description": "...", "keywords": ["..."]}}""",
model="minicpm-v4.5:8b", host=OLLAMA_GAMINGPC, temperature=0.3)
model="qwen3.8:latest", host=OLLAMA_SHADOW, temperature=0.3)
# Agent 5: Real Fact Check (web-verified)
factcheck = real_fact_check(edited, topic_title)
@@ -942,7 +944,7 @@ ARTICLE:
{edited}
Return the expanded article in full Markdown. No JSON wrapper.""",
model="minicpm-v4.5:8b", host=OLLAMA_GAMINGPC, temperature=0.5, max_tokens=8192)
model="qwen3.8:latest", host=OLLAMA_SHADOW, temperature=0.5, max_tokens=8192)
passed, issues = quality_gate(edited, topic_title, vertical)
if not passed: