v6: Cloud fallback, quality gate, format diversity, real fact-check, scroll depth, email capture, dual-model research, vision-verified images, network footer
This commit is contained in:
drjones
@@ -142,61 +142,81 @@ def init_db():
|
||||
|
||||
|
||||
# ─── LLM Helpers ───────────────────────────────────────────────────
|
||||
def ollama_chat(prompt: str, model: str = "qwen3.5:4b", host: str = OLLAMA_MACBOOK,
|
||||
system: str = "", temperature: float = 0.7, max_tokens: int = 4096) -> str:
|
||||
"""Call Ollama chat API. Falls back to GamingPC if MacBook fails."""
|
||||
DEEPSEEK_API_KEY = os.environ.get("DEEPSEEK_API_KEY", "")
|
||||
DEEPSEEK_API_URL = "https://api.deepseek.com/v1/chat/completions"
|
||||
|
||||
def _call_deepseek(prompt: str, model: str = "deepseek-chat", system: str = "",
|
||||
temperature: float = 0.7, max_tokens: int = 4096) -> str:
|
||||
"""Call DeepSeek cloud API as fallback."""
|
||||
if not DEEPSEEK_API_KEY:
|
||||
raise RuntimeError("No DEEPSEEK_API_KEY set")
|
||||
payload = {
|
||||
"model": model,
|
||||
"messages": [],
|
||||
"stream": False,
|
||||
"options": {
|
||||
"temperature": temperature,
|
||||
"num_predict": max_tokens,
|
||||
}
|
||||
"temperature": temperature,
|
||||
"max_tokens": max_tokens,
|
||||
}
|
||||
if system:
|
||||
payload["messages"].append({"role": "system", "content": system})
|
||||
payload["messages"].append({"role": "user", "content": prompt})
|
||||
|
||||
r = requests.post(DEEPSEEK_API_URL, json=payload,
|
||||
headers={"Authorization": f"Bearer {DEEPSEEK_API_KEY}",
|
||||
"Content-Type": "application/json"},
|
||||
timeout=120)
|
||||
if r.status_code == 200:
|
||||
return r.json()["choices"][0]["message"]["content"]
|
||||
raise RuntimeError(f"DeepSeek API error {r.status_code}: {r.text[:200]}")
|
||||
|
||||
|
||||
def llm_chat(prompt: str, model: str = "qwen3.5:4b", host: str = OLLAMA_MACBOOK,
|
||||
system: str = "", temperature: float = 0.7, max_tokens: int = 4096,
|
||||
retries: int = 3) -> str:
|
||||
"""Call LLM with Ollama → DeepSeek fallback, with retries."""
|
||||
payload = {
|
||||
"model": model, "messages": [], "stream": False,
|
||||
"options": {"temperature": temperature, "num_predict": max_tokens}
|
||||
}
|
||||
if system:
|
||||
payload["messages"].append({"role": "system", "content": system})
|
||||
payload["messages"].append({"role": "user", "content": prompt})
|
||||
|
||||
hosts = [host]
|
||||
if host == OLLAMA_MACBOOK:
|
||||
hosts.append(OLLAMA_GAMINGPC)
|
||||
# Also try the other if not already in list
|
||||
if OLLAMA_GAMINGPC not in hosts:
|
||||
hosts.append(OLLAMA_GAMINGPC)
|
||||
if OLLAMA_MACBOOK not in hosts:
|
||||
hosts.append(OLLAMA_MACBOOK)
|
||||
# Try Ollama hosts first
|
||||
hosts = list(dict.fromkeys([host, OLLAMA_MACBOOK, 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),
|
||||
proxies={"http": None, "https": None})
|
||||
if r.status_code == 200:
|
||||
result = r.json()
|
||||
if "message" in result:
|
||||
return result["message"]["content"]
|
||||
if "error" in result:
|
||||
log.warning(f"Ollama {h} error: {result['error']}")
|
||||
continue
|
||||
except Exception as e:
|
||||
log.warning(f"Ollama {h} attempt {attempt+1} failed: {e}")
|
||||
continue
|
||||
if attempt < retries - 1:
|
||||
time.sleep(2 ** attempt)
|
||||
|
||||
for h in hosts:
|
||||
# Cloud fallback
|
||||
if DEEPSEEK_API_KEY:
|
||||
log.info("All Ollama hosts failed — falling back to DeepSeek cloud")
|
||||
try:
|
||||
r = requests.post(f"{h}/api/chat", json=payload, timeout=300,
|
||||
proxies={"http": None, "https": None})
|
||||
if r.status_code == 200:
|
||||
result = r.json()
|
||||
if "message" in result:
|
||||
return result["message"]["content"]
|
||||
if "error" in result:
|
||||
log.warning(f"Ollama {h} error: {result['error']}")
|
||||
continue
|
||||
return _call_deepseek(prompt, system=system, temperature=temperature, max_tokens=max_tokens)
|
||||
except Exception as e:
|
||||
log.warning(f"Ollama {h} failed: {e}")
|
||||
continue
|
||||
log.error(f"DeepSeek fallback also failed: {e}")
|
||||
|
||||
# Fallback: use active cloud LLM (DeepSeek) via Hermes tools
|
||||
log.warning("All Ollama hosts failed/saturated — falling back to cloud LLM")
|
||||
raise RuntimeError(
|
||||
f"All Ollama hosts failed for model {model}. "
|
||||
"Local LLMs are saturated (likely by Kalshi bots). "
|
||||
"Retry when load is lower or add cloud fallback API key."
|
||||
)
|
||||
raise RuntimeError(f"All LLM hosts failed for model {model}")
|
||||
|
||||
|
||||
def ollama_json(prompt: str, model: str = "qwen3.5:4b", host: str = OLLAMA_MACBOOK,
|
||||
system: str = "You are a JSON-only API. Always respond with valid JSON. No markdown, no explanation.",
|
||||
temperature: float = 0.3) -> dict:
|
||||
"""Call Ollama and parse JSON response."""
|
||||
raw = ollama_chat(prompt, model=model, host=host, system=system, temperature=temperature)
|
||||
# Strip markdown code fences if present
|
||||
def llm_json(prompt: str, model: str = "qwen3.5:4b", host: str = OLLAMA_MACBOOK,
|
||||
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."""
|
||||
raw = llm_chat(prompt, model=model, host=host, system=system, temperature=temperature)
|
||||
raw = raw.strip()
|
||||
if raw.startswith("```"):
|
||||
lines = raw.split("\n")
|
||||
@@ -204,6 +224,67 @@ def ollama_json(prompt: str, model: str = "qwen3.5:4b", host: str = OLLAMA_MACBO
|
||||
return json.loads(raw)
|
||||
|
||||
|
||||
def dual_llm_research(prompt: str, system: str = "") -> tuple[str, dict]:
|
||||
"""Run research on two models in parallel. Returns (merged_output, disagreement_report)."""
|
||||
import concurrent.futures
|
||||
|
||||
def call_ornith():
|
||||
return llm_chat(prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
|
||||
system=system, temperature=0.3, max_tokens=4096)
|
||||
|
||||
def call_qwen():
|
||||
return llm_chat(prompt, model="qwen3.5:4b", host=OLLAMA_MACBOOK,
|
||||
system=system, temperature=0.3, max_tokens=2048)
|
||||
|
||||
with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
|
||||
future_ornith = executor.submit(call_ornith)
|
||||
future_qwen = executor.submit(call_qwen)
|
||||
|
||||
try:
|
||||
ornith_result = future_ornith.result(timeout=180)
|
||||
except Exception as e:
|
||||
log.warning(f"Ornith research failed: {e}")
|
||||
ornith_result = None
|
||||
|
||||
try:
|
||||
qwen_result = future_qwen.result(timeout=60)
|
||||
except Exception as e:
|
||||
log.warning(f"Qwen research failed: {e}")
|
||||
qwen_result = None
|
||||
|
||||
# Merge: ornith leads, qwen fills gaps
|
||||
if ornith_result:
|
||||
if qwen_result:
|
||||
# Quick disagreement check
|
||||
disagreements = _check_disagreements(ornith_result, qwen_result)
|
||||
return ornith_result, disagreements
|
||||
return ornith_result, {}
|
||||
elif qwen_result:
|
||||
return qwen_result, {}
|
||||
else:
|
||||
raise RuntimeError("Both research models failed")
|
||||
|
||||
|
||||
def _check_disagreements(text1: str, text2: str) -> dict:
|
||||
"""Quick check for factual disagreements between two outputs."""
|
||||
# Lightweight: extract capitalized entities and numbers, compare
|
||||
import re
|
||||
entities1 = set(re.findall(r'[A-Z][a-z]+(?:\s[A-Z][a-z]+)*', text1))
|
||||
entities2 = set(re.findall(r'[A-Z][a-z]+(?:\s[A-Z][a-z]+)*', text2))
|
||||
only_in_1 = entities1 - entities2
|
||||
only_in_2 = entities2 - entities1
|
||||
numbers1 = set(re.findall(r'\d+(?:\.\d+)?%?', text1))
|
||||
numbers2 = set(re.findall(r'\d+(?:\.\d+)?%?', text2))
|
||||
num_diff = numbers1.symmetric_difference(numbers2)
|
||||
|
||||
return {
|
||||
"disagreed": len(only_in_1) > 5 or len(only_in_2) > 5,
|
||||
"entities_only_in_first": list(only_in_1)[:10],
|
||||
"entities_only_in_second": list(only_in_2)[:10],
|
||||
"number_mismatches": list(num_diff)[:10],
|
||||
}
|
||||
|
||||
|
||||
# ─── Trend Discovery ────────────────────────────────────────────────
|
||||
TREND_SOURCES = [
|
||||
{"name": "Hacker News", "url": "https://hacker-news.firebaseio.com/v0/topstories.json"},
|
||||
@@ -348,7 +429,7 @@ 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 = ollama_json(prompt, model="qwen3.5:4b", temperature=0.8)
|
||||
result = llm_json(prompt, model="qwen3.5:4b", temperature=0.8)
|
||||
if isinstance(result, list):
|
||||
return result
|
||||
return list(result.values())[0] if result else []
|
||||
@@ -400,7 +481,7 @@ Vertical assignment rules:
|
||||
Respond with a JSON array of objects. No markdown, no explanation."""
|
||||
|
||||
try:
|
||||
result = ollama_json(prompt, model="qwen3.5:4b", temperature=0.3)
|
||||
result = llm_json(prompt, model="qwen3.5:4b", temperature=0.3)
|
||||
if isinstance(result, list):
|
||||
# Apply algorithmic boost on top of LLM scores
|
||||
return _apply_learning_boost(result, learning_insights)
|
||||
@@ -549,11 +630,11 @@ Extract and return as JSON:
|
||||
Be accurate. Cite real sources. No hallucinations. Respond with ONLY valid JSON."""
|
||||
|
||||
try:
|
||||
result = ollama_json(research_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
|
||||
result = llm_json(research_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
|
||||
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 = ollama_json(research_prompt, model="qwen3.5:4b",
|
||||
result = llm_json(research_prompt, model="qwen3.5:4b",
|
||||
system="You are an expert research analyst. Be accurate and honest.")
|
||||
|
||||
# Store knowledge package
|
||||
@@ -605,167 +686,296 @@ def _web_search_sources(topic: str) -> list[dict]:
|
||||
return sources
|
||||
|
||||
|
||||
# ─── Article Formats ──────────────────────────────────────────────
|
||||
ARTICLE_FORMATS = [
|
||||
{"name": "explainer", "weight": 35, "target_words": "1500-3000",
|
||||
"desc": "Comprehensive deep-dive explainer with sections, examples, and FAQ"},
|
||||
{"name": "listicle", "weight": 20, "target_words": "1200-2000",
|
||||
"desc": "Numbered list format: '7 Ways to...', '5 Reasons Why...', etc"},
|
||||
{"name": "quick-tip", "weight": 10, "target_words": "400-800",
|
||||
"desc": "Short, focused practical tip or trick. One clear takeaway"},
|
||||
{"name": "deep-dive", "weight": 15, "target_words": "2500-4000",
|
||||
"desc": "Exhaustive technical deep-dive with code, data, and analysis"},
|
||||
{"name": "comparison", "weight": 10, "target_words": "1500-2500",
|
||||
"desc": "Head-to-head comparison: X vs Y with pros/cons and verdict"},
|
||||
{"name": "news-roundup", "weight": 10, "target_words": "800-1500",
|
||||
"desc": "Weekly-style roundup of latest developments in a topic area"},
|
||||
]
|
||||
|
||||
def _pick_format() -> dict:
|
||||
"""Randomly select an article format weighted by preference."""
|
||||
import random
|
||||
total = sum(f["weight"] for f in ARTICLE_FORMATS)
|
||||
r = random.uniform(0, total)
|
||||
cumulative = 0
|
||||
for fmt in ARTICLE_FORMATS:
|
||||
cumulative += fmt["weight"]
|
||||
if r <= cumulative:
|
||||
return fmt
|
||||
return ARTICLE_FORMATS[0]
|
||||
|
||||
# ─── Quality Gate ──────────────────────────────────────────────────
|
||||
AI_CLICHES = [
|
||||
"delve", "unleash", "game-changer", "in today's world", "it's important to note",
|
||||
"revolutionary", "groundbreaking", "game changing", "cutting-edge",
|
||||
"in the fast-paced world", "a testament to", "it is worth noting",
|
||||
"paradigm shift", "in this digital age", "unprecedented",
|
||||
]
|
||||
|
||||
def quality_gate(article_text: str, title: str, vertical: str) -> tuple[bool, list[str]]:
|
||||
"""Pre-publish quality checks. Returns (passed, issues)."""
|
||||
issues = []
|
||||
wc = len(article_text.split())
|
||||
|
||||
# Word count check
|
||||
if wc < 400:
|
||||
issues.append(f"Too short: {wc} words (minimum 400)")
|
||||
|
||||
# AI cliché check
|
||||
cliches_found = [c for c in AI_CLICHES if c.lower() in article_text.lower()]
|
||||
if cliches_found:
|
||||
issues.append(f"AI clichés: {', '.join(cliches_found[:5])}")
|
||||
|
||||
# Basic readability: check for very long sentences (>50 words)
|
||||
long_sentences = [s for s in article_text.replace('!', '.').replace('?', '.').split('.')
|
||||
if len(s.split()) > 50]
|
||||
if len(long_sentences) > 5:
|
||||
issues.append(f"{len(long_sentences)} sentences exceed 50 words — hard to read")
|
||||
|
||||
# Empty content check
|
||||
if not article_text.strip() or len(article_text) < 200:
|
||||
issues.append("Article appears empty or truncated")
|
||||
|
||||
return len(issues) == 0, issues
|
||||
|
||||
|
||||
# ─── Real Fact Checker ────────────────────────────────────────────
|
||||
def real_fact_check(article_text: str, topic_title: str) -> dict:
|
||||
"""Verify factual claims by searching the web."""
|
||||
claims = []
|
||||
# Extract claims: sentences with numbers, percentages, or specific facts
|
||||
import re
|
||||
for sentence in article_text.split('.')[:30]: # First 30 sentences
|
||||
s = sentence.strip()
|
||||
if not s:
|
||||
continue
|
||||
has_stat = bool(re.search(r'\d+%|\d+\s(?:million|billion|thousand)|according to|study|research|found that', s, re.I))
|
||||
if has_stat and len(s) > 40:
|
||||
claims.append(s[:300])
|
||||
|
||||
if len(claims) < 2:
|
||||
return {"verified": True, "checked": 0, "issues": []}
|
||||
|
||||
# Search web for each claim
|
||||
issues = []
|
||||
verified_count = 0
|
||||
for claim in claims[:5]: # Check up to 5 claims
|
||||
try:
|
||||
search_query = claim[:150]
|
||||
r = requests.get(
|
||||
f"https://api.duckduckgo.com/?q={requests.utils.quote(search_query)}&format=json&no_html=1",
|
||||
timeout=10, headers={"User-Agent": "AutoPublisher/2.0"}
|
||||
)
|
||||
if r.status_code == 200:
|
||||
data = r.json()
|
||||
abstract = data.get("AbstractText", "") or data.get("Abstract", "")
|
||||
if abstract and len(abstract) > 30:
|
||||
verified_count += 1
|
||||
else:
|
||||
issues.append(f"Could not verify: '{claim[:100]}...'")
|
||||
except Exception:
|
||||
pass # Web search failed — not critical enough to block
|
||||
|
||||
return {
|
||||
"verified": len(issues) == 0,
|
||||
"checked": len(claims[:5]),
|
||||
"verified_count": verified_count,
|
||||
"issues": issues,
|
||||
}
|
||||
|
||||
|
||||
# ─── Image Generator Hook ──────────────────────────────────────────
|
||||
def generate_article_image(title: str, vertical: str) -> str | None:
|
||||
"""Generate a hero image for an article via FAL.ai, verify with vision model. Returns URL or None."""
|
||||
# Build a prompt that captures the article's essence
|
||||
prompt = f"""Dark atmospheric illustration for an article titled "{title}".
|
||||
Vertical: {vertical}. Clean, minimal, modern. No text. Wide cinematic composition.
|
||||
Dark background matching the site's aesthetic. Abstract but relevant to the topic. Premium quality."""
|
||||
|
||||
try:
|
||||
# Call FAL via Nous subscription
|
||||
r = requests.post("http://localhost:5106/api/generate-image",
|
||||
json={"prompt": prompt, "aspect_ratio": "landscape"},
|
||||
timeout=30)
|
||||
if r.status_code != 200:
|
||||
log.info("Image gen not available — using site hero fallback")
|
||||
return f"/assets/hero.png"
|
||||
|
||||
image_url = r.json().get("image_url", "")
|
||||
if not image_url:
|
||||
return f"/assets/hero.png"
|
||||
|
||||
# Verify image with local vision model
|
||||
try:
|
||||
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,
|
||||
system="You are an image quality reviewer. Be strict but fair.",
|
||||
temperature=0.1,
|
||||
max_tokens=50,
|
||||
)
|
||||
if "FAIL" in verify:
|
||||
log.warning(f"Image verification failed: {verify}")
|
||||
return f"/assets/hero.png"
|
||||
log.info(f"Image verified by vision model: {verify}")
|
||||
except Exception as e:
|
||||
log.warning(f"Vision model check skipped: {e}")
|
||||
|
||||
return image_url
|
||||
except Exception as e:
|
||||
log.warning(f"Image generation failed: {e}")
|
||||
return f"/assets/hero.png"
|
||||
|
||||
|
||||
# ─── Writing Pipeline ──────────────────────────────────────────────
|
||||
def write_article(topic_id: int, topic_title: str, vertical: str,
|
||||
knowledge_package: dict) -> dict:
|
||||
"""Multi-agent writing pipeline: outline → draft → SEO → edit → fact-check."""
|
||||
knowledge_package: dict) -> dict | None:
|
||||
"""Multi-agent writing pipeline with format diversity, quality gate, and fact-check."""
|
||||
log.info(f"Writing article for topic #{topic_id}: {topic_title}")
|
||||
|
||||
|
||||
kp_json = json.dumps(knowledge_package, indent=2)
|
||||
|
||||
# Agent 1: Outline
|
||||
outline_prompt = f"""Create a detailed article outline for:
|
||||
fmt = _pick_format()
|
||||
log.info(f" Format: {fmt['name']} ({fmt['target_words']} words)")
|
||||
|
||||
# Agent 1: Outline (adapted to format)
|
||||
outline_prompt = f"""Create a detailed article outline for a {fmt['name']} format article.
|
||||
|
||||
TITLE: {topic_title}
|
||||
VERTICAL: {vertical}
|
||||
FORMAT: {fmt['name']} — {fmt['desc']}
|
||||
TARGET: {fmt['target_words']} words
|
||||
|
||||
KNOWLEDGE PACKAGE:
|
||||
{kp_json}
|
||||
|
||||
Generate an outline with:
|
||||
- Introduction hook
|
||||
- 5-8 major sections with subsections
|
||||
- Key takeaways
|
||||
- FAQ section topics
|
||||
- Call-to-action
|
||||
|
||||
Generate an outline appropriate for this format.
|
||||
Respond with JSON:
|
||||
{{"sections": [{{"heading": "...", "subsections": ["..."]}}, ...], "faq_questions": ["..."], "cta": "..."}}"""
|
||||
|
||||
outline = ollama_json(outline_prompt, model="qwen3.5:4b", temperature=0.5)
|
||||
|
||||
# Agent 2: Technical Writer (ornith for quality)
|
||||
draft_prompt = f"""Write a comprehensive, authoritative article.
|
||||
outline = llm_json(outline_prompt, model="qwen3.5:4b", temperature=0.5)
|
||||
|
||||
# Agent 2: Draft with format guidance
|
||||
draft_prompt = f"""Write a {fmt['name']} format article.
|
||||
|
||||
TITLE: {topic_title}
|
||||
VERTICAL: {vertical}
|
||||
FORMAT: {fmt['name']} — {fmt['desc']}
|
||||
TARGET: {fmt['target_words']} words
|
||||
OUTLINE: {json.dumps(outline)}
|
||||
FACTS: {json.dumps(knowledge_package.get('facts', []))}
|
||||
STATS: {json.dumps(knowledge_package.get('stats', []))}
|
||||
DEFINITIONS: {json.dumps(knowledge_package.get('definitions', []))}
|
||||
EXAMPLES: {json.dumps(knowledge_package.get('examples', []))}
|
||||
CITATIONS: {json.dumps(knowledge_package.get('citations', []))}
|
||||
|
||||
Write the full article in clean Markdown. Include:
|
||||
Requirements:
|
||||
- Match the {fmt['name']} format naturally
|
||||
- Engaging introduction that hooks the reader
|
||||
- Well-structured sections following the outline
|
||||
- Code blocks where relevant (for tech/linux)
|
||||
- Pull quotes from key stats
|
||||
- Real, specific details — not generic filler
|
||||
- "Key Takeaway" boxes (use > blockquotes)
|
||||
- FAQ section at the end
|
||||
- Sources/citations section
|
||||
|
||||
Target: 1500-3000 words. Use a clear, authoritative but conversational tone.
|
||||
DO NOT use AI clichés ("delve", "unleash", "game-changer", "in today's world").
|
||||
Write like an expert explaining to an intelligent peer.
|
||||
- FAQ section at the end where relevant
|
||||
- Use a clear, authoritative but conversational tone
|
||||
- DO NOT use AI clichés (delve, unleash, game-changer, in today's world, revolutionary, groundbreaking, cutting-edge, unprecedented, paradigm shift)
|
||||
|
||||
Respond with the FULL Markdown article. No JSON wrapper."""
|
||||
|
||||
draft = ollama_chat(draft_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
|
||||
system="You are an expert technical writer. Write clear, accurate, engaging content. No AI clichés. No fluff.",
|
||||
temperature=0.7, max_tokens=8192)
|
||||
|
||||
# Agent 3: Copy Editor (qwen, fast)
|
||||
edit_prompt = f"""Edit and improve this article. Fix:
|
||||
- Grammar and spelling
|
||||
- Awkward phrasing
|
||||
- Repetition
|
||||
- Clarity issues
|
||||
- Add transitions between sections
|
||||
- Ensure consistent tone
|
||||
- Break up overly long paragraphs
|
||||
draft = llm_chat(draft_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
|
||||
system="You are an expert writer. Write clear, accurate, engaging content. No AI clichés. No fluff.",
|
||||
temperature=0.75, max_tokens=8192)
|
||||
|
||||
# Agent 3: Copy Editor
|
||||
edited = llm_chat(
|
||||
f"""Edit and improve this article. Fix grammar, awkward phrasing, repetition. Add transitions. Break up long paragraphs. Ensure consistent tone.
|
||||
|
||||
ARTICLE:
|
||||
{draft}
|
||||
|
||||
Return the edited article in full Markdown. No JSON wrapper."""
|
||||
|
||||
edited = ollama_chat(edit_prompt, model="qwen3.5:4b", temperature=0.3, max_tokens=8192)
|
||||
|
||||
# Agent 4: SEO Optimization
|
||||
seo_prompt = f"""Optimize this article for SEO. Generate:
|
||||
|
||||
1. SEO title (55-65 chars, include primary keyword)
|
||||
2. Meta description (150-160 chars, compelling)
|
||||
3. Suggested internal links (related topics from same vertical)
|
||||
4. Tags/keywords (5-10)
|
||||
|
||||
ARTICLE TITLE: {topic_title}
|
||||
VERTICAL: {vertical}
|
||||
Return the edited article in full Markdown. No JSON wrapper.""",
|
||||
model="qwen3.5:4b", 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": ["..."], "internal_links": [{{"text": "...", "slug": "..."}}]}}"""
|
||||
|
||||
seo = ollama_json(seo_prompt, model="qwen3.5:4b", temperature=0.3)
|
||||
|
||||
# Agent 5: Fact Check (ornith)
|
||||
factcheck_prompt = f"""Fact check this article. Verify:
|
||||
1. Are the statistics accurate and properly sourced?
|
||||
2. Are any claims unsubstantiated?
|
||||
3. Are technical details correct?
|
||||
4. Are dates and timelines accurate?
|
||||
5. Is anything overstated or misleading?
|
||||
|
||||
ARTICLE:
|
||||
{edited[:4000]}
|
||||
|
||||
FACTS USED:
|
||||
{json.dumps(knowledge_package.get('facts', []))}
|
||||
|
||||
Respond with JSON:
|
||||
{{"passed": true/false, "issues": ["issue 1", ...], "corrections": [{{"original": "...", "corrected": "..."}}]}}"""
|
||||
|
||||
factcheck = ollama_json(factcheck_prompt, model="ornith:latest", host=OLLAMA_GAMINGPC,
|
||||
system="You are a strict fact-checker. Flag everything questionable. Be conservative — if unsure, flag it.",
|
||||
temperature=0.1)
|
||||
|
||||
# If fact check found issues, apply corrections
|
||||
if not factcheck.get("passed", True):
|
||||
corrections = factcheck.get("corrections", [])
|
||||
if corrections:
|
||||
fix_prompt = f"""Apply these corrections to the article:
|
||||
|
||||
{json.dumps(corrections, indent=2)}
|
||||
Respond with JSON: {{"seo_title": "...", "seo_description": "...", "keywords": ["..."]}}""",
|
||||
model="qwen3.5:4b", temperature=0.3)
|
||||
|
||||
# Agent 5: Real Fact Check (web-verified)
|
||||
factcheck = real_fact_check(edited, topic_title)
|
||||
if not factcheck["verified"]:
|
||||
log.warning(f" Fact-check issues: {factcheck['issues']}")
|
||||
|
||||
# Agent 6: Quality Gate
|
||||
passed, issues = quality_gate(edited, topic_title, vertical)
|
||||
if not passed:
|
||||
log.warning(f" Quality gate FAILED: {issues}")
|
||||
# Try to fix common issues
|
||||
if any("Too short" in i for i in issues):
|
||||
# Expand the article
|
||||
edited = llm_chat(
|
||||
f"""This article is too short. Expand it with more detail, examples, and depth. Keep the same tone and structure.
|
||||
|
||||
ARTICLE:
|
||||
{edited}
|
||||
|
||||
Return the corrected article in full Markdown. No JSON wrapper."""
|
||||
edited = ollama_chat(fix_prompt, model="qwen3.5:4b", temperature=0.2)
|
||||
|
||||
Return the expanded article in full Markdown. No JSON wrapper.""",
|
||||
model="qwen3.5:4b", temperature=0.5, max_tokens=8192)
|
||||
passed, issues = quality_gate(edited, topic_title, vertical)
|
||||
|
||||
if not passed:
|
||||
log.error(f" Quality gate STILL failing after fix: {issues}")
|
||||
# Store as draft, don't publish
|
||||
db = init_db()
|
||||
db.execute("UPDATE topics SET status = 'quality_failed' WHERE id = ?", (topic_id,))
|
||||
db.commit()
|
||||
db.close()
|
||||
return None
|
||||
|
||||
# Agent 7: Generate article image
|
||||
og_image = generate_article_image(topic_title, vertical)
|
||||
|
||||
# Calculate stats
|
||||
word_count = len(edited.split())
|
||||
reading_time = max(1, word_count // 200)
|
||||
slug = topic_title.lower().strip()[:80]
|
||||
slug = "".join(c if c.isalnum() or c in "- " else "" for c in slug)
|
||||
slug = slug.replace(" ", "-").strip("-")
|
||||
|
||||
|
||||
# Store article
|
||||
db = init_db()
|
||||
db.execute("""
|
||||
INSERT OR REPLACE INTO articles (topic_id, vertical, title, slug, content_md,
|
||||
seo_title, seo_description, word_count, reading_time_minutes, status)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, 'draft')
|
||||
seo_title, seo_description, og_image, word_count, reading_time_minutes, status)
|
||||
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 'draft')
|
||||
""", (topic_id, vertical, topic_title, slug, edited,
|
||||
seo.get("seo_title", topic_title[:65]),
|
||||
seo.get("seo_description", ""),
|
||||
og_image or "",
|
||||
word_count, reading_time))
|
||||
db.execute("UPDATE topics SET status = 'written', article_id = last_insert_rowid() WHERE id = ?",
|
||||
(topic_id,))
|
||||
db.commit()
|
||||
db.close()
|
||||
|
||||
log.info(f"Article written for #{topic_id}: {word_count} words, {reading_time}min read")
|
||||
|
||||
log.info(f"Article written for #{topic_id}: {word_count}w, {reading_time}min, format={fmt['name']}, "
|
||||
f"fact_checked={factcheck['verified_count']}/{factcheck['checked']}, quality=OK")
|
||||
return {
|
||||
"topic_id": topic_id,
|
||||
"title": topic_title,
|
||||
"slug": slug,
|
||||
"content": edited,
|
||||
"seo": seo,
|
||||
"word_count": word_count,
|
||||
"reading_time": reading_time,
|
||||
"factcheck_passed": factcheck.get("passed", True),
|
||||
"topic_id": topic_id, "title": topic_title, "slug": slug,
|
||||
"content_md": edited, "content": edited, "seo": seo,
|
||||
"word_count": word_count, "reading_time": reading_time,
|
||||
"og_image": og_image, "factcheck": factcheck, "format": fmt['name'],
|
||||
}
|
||||
|
||||
|
||||
|
||||
@@ -234,6 +234,14 @@ def init_db():
|
||||
CREATE INDEX IF NOT EXISTS idx_pageviews_created ON pageviews(created_at);
|
||||
CREATE INDEX IF NOT EXISTS idx_articles_published ON articles(published_at);
|
||||
CREATE INDEX IF NOT EXISTS idx_articles_slug ON articles(slug);
|
||||
|
||||
CREATE TABLE IF NOT EXISTS subscribers (
|
||||
id INTEGER PRIMARY KEY AUTOINCREMENT,
|
||||
email TEXT UNIQUE NOT NULL,
|
||||
vertical TEXT DEFAULT '',
|
||||
confirmed INTEGER DEFAULT 0,
|
||||
created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
|
||||
);
|
||||
""")
|
||||
|
||||
|
||||
@@ -601,6 +609,22 @@ def api_learn():
|
||||
})
|
||||
|
||||
|
||||
@app.route("/api/subscribe", methods=["POST"])
|
||||
def api_subscribe():
|
||||
"""Email newsletter signup."""
|
||||
email = (request.json or {}).get("email", "").strip().lower()
|
||||
if not email or "@" not in email:
|
||||
return jsonify({"error": "invalid email"}), 400
|
||||
db = get_db()
|
||||
try:
|
||||
db.execute("INSERT OR IGNORE INTO subscribers (email, vertical) VALUES (?, ?)",
|
||||
(email, VERTICAL))
|
||||
db.commit()
|
||||
return jsonify({"status": "subscribed"})
|
||||
except Exception as e:
|
||||
return jsonify({"error": str(e)}), 500
|
||||
|
||||
|
||||
@app.route("/health")
|
||||
def health():
|
||||
db = get_db()
|
||||
@@ -1024,6 +1048,13 @@ ARTICLE_TEMPLATE = """<!DOCTYPE html>
|
||||
.network-link span{display:block;color:var(--text-muted);font-size:0.7rem;font-weight:400;margin-top:0.1rem}
|
||||
.footer-bottom{display:flex;justify-content:space-between;align-items:center;flex-wrap:wrap;gap:1rem;color:var(--text-muted);font-size:0.8rem}
|
||||
.footer-bottom a{color:var(--text-muted);text-decoration:none}
|
||||
|
||||
/* Newsletter */
|
||||
.newsletter-box{background:var(--card-bg);border:1px solid var(--border);border-radius:12px;padding:1.25rem;margin-top:2rem;text-align:center}
|
||||
.newsletter-box h4{font-family:var(--font-heading);font-size:1rem;margin-bottom:0.75rem}
|
||||
.subscribe-form{display:flex;gap:0.5rem;max-width:400px;margin:0 auto}
|
||||
.subscribe-form input{flex:1;padding:0.6rem 0.75rem;background:var(--bg);border:1px solid var(--border);border-radius:6px;color:var(--text);font-size:0.9rem}
|
||||
.subscribe-form button{background:var(--gradient);color:white;border:none;padding:0.6rem 1.25rem;border-radius:6px;cursor:pointer;font-weight:600;font-size:0.9rem}
|
||||
</style>
|
||||
</head>
|
||||
<body>
|
||||
@@ -1077,11 +1108,21 @@ ARTICLE_TEMPLATE = """<!DOCTYPE html>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="newsletter-box">
|
||||
<h4>📬 Get new articles by email</h4>
|
||||
<form class="subscribe-form" onsubmit="subscribe(event)">
|
||||
<input type="email" id="sub-email" placeholder="your@email.com" required>
|
||||
<button type="submit">Subscribe</button>
|
||||
</form>
|
||||
<div id="sub-msg" style="margin-top:0.5rem;font-size:0.8rem;display:none"></div>
|
||||
<p style="font-size:0.7rem;color:var(--text-muted);margin-top:0.5rem">No spam. Just new articles from {{ name }}.</p>
|
||||
</div>
|
||||
|
||||
{% if related %}
|
||||
<section class="related">
|
||||
<h2>Continue Reading</h2>
|
||||
<div class="related-grid">
|
||||
{% for r in related %}
|
||||
{% for r in related[:3] %}
|
||||
<a href="/articles/{{ r.slug }}" class="related-card">
|
||||
<h4>{{ r.title[:60] }}</h4>
|
||||
<div class="meta">{{ r.reading_time }} min · {{ r.published_at[:10] }}</div>
|
||||
@@ -1139,6 +1180,36 @@ ARTICLE_TEMPLATE = """<!DOCTYPE html>
|
||||
toc.appendChild(li);
|
||||
});
|
||||
})();
|
||||
|
||||
// Scroll depth tracking
|
||||
(function(){
|
||||
var fired = {25:false,50:false,75:false,100:false};
|
||||
window.addEventListener('scroll', function(){
|
||||
var pct = Math.round((window.scrollY / (document.documentElement.scrollHeight - window.innerHeight)) * 100);
|
||||
[25,50,75,100].forEach(function(threshold){
|
||||
if(pct >= threshold && !fired[threshold]){
|
||||
fired[threshold] = true;
|
||||
new Image().src = '/a/ping?p=/articles/{{ article.slug }}&d=' + threshold;
|
||||
}
|
||||
});
|
||||
});
|
||||
// Time on page ping every 30s
|
||||
setInterval(function(){
|
||||
new Image().src = '/a/ping?p=/articles/{{ article.slug }}&t=1';
|
||||
}, 30000);
|
||||
})();
|
||||
|
||||
function subscribe(e){
|
||||
e.preventDefault();
|
||||
var email = document.getElementById('sub-email').value;
|
||||
var msg = document.getElementById('sub-msg');
|
||||
fetch('/api/subscribe', {method:'POST',headers:{'Content-Type':'application/json'},body:JSON.stringify({email:email})})
|
||||
.then(function(r){ return r.json(); })
|
||||
.then(function(d){
|
||||
msg.style.display = 'block';
|
||||
msg.textContent = d.status === 'subscribed' ? '✅ Subscribed! Welcome.' : '❌ ' + (d.error || 'Error');
|
||||
});
|
||||
}
|
||||
</script>
|
||||
<img src="/a/ping?p=/articles/{{ article.slug }}" alt="" width="1" height="1" style="display:none">
|
||||
</body>
|
||||
|
||||
Reference in New Issue
Block a user