Miselenstralium / agent.py
acecalisto3's picture
Update agent.py
57458ec verified
"""
Enhanced Viral Content Agent
- Deterministic, testable, dependency-light
- Action loop with tool allow‑list and guarded parsing
- Pluggable LLM backends (Hugging Face Inference API, OpenAI, generic HTTP JSON API) with graceful fallback
- Research tool with real HTTP search (DuckDuckGo HTML) + Wikipedia summary fallback; offline synthetic fallback retained
- JSONL logging and reproducible runs via seed
Runtime targets: Python 3.9+
"""
from __future__ import annotations
import os
import re
import io
import json
import time
import uuid
import math
import random
import logging
import contextlib
from dataclasses import dataclass, field
from datetime import datetime, timezone
from typing import Any, Dict, List, Optional, Tuple, Iterable
import requests
# ---------------------------
# Logging
# ---------------------------
LOGGER_NAME = "viral_agent"
logger = logging.getLogger(LOGGER_NAME)
if not logger.handlers:
level = os.getenv("AGENT_LOG_LEVEL", "INFO").upper()
logging.basicConfig(level=getattr(logging, level, logging.INFO), format="%(asctime)s %(levelname)s | %(message)s")
# ---------------------------
# Prompts (kept concise; multi‑line strings)
# ---------------------------
PREFIX = (
"You are an Advanced Viral Content Generator with self‑research and self‑improvement capabilities.\n"
"Tools: GENERATE_IDEA, RESEARCH, GENERATE_CONTENT, SELF_EVALUATE, IMPROVE_CONTENT, FORMAT_CONTENT, PUBLISH, COMPLETE.\n"
"Trigger using lines: action: <TOOL> and action_input=<TEXT>.\n"
)
IDEA_GENERATOR_PROMPT = (
"Generate one viral content idea. Consider trending topics, underserved niches, controversy, practical value, and emotion.\n"
"Return a single concise title. Topic: {topic}. History: {history}"
)
RESEARCH_PROMPT = (
"You are researching: {topic}. Summarize key facts with bullet points. Include stats with sources when available."
)
CONTENT_PROMPT = (
"Create {format_type} content about: {topic}. Use the following research notes: {research}.\n"
"Hook, sections with headings, and a clear wrap‑up. Keep it factual and concise."
)
EVALUATE_PROMPT = (
"Evaluate content quality and viral potential from 1‑10 for engagement, accuracy, originality, emotion, readability, and headline strength.\n"
"Return compact JSON with fields per_criterion and overall plus three specific improvements. Content: {content}"
)
IMPROVE_PROMPT = (
"Improve the content using this feedback: {feedback}. Strengthen hook, structure, and specificity. Return the full revised content. Content: {content}"
)
FORMAT_PROMPT = (
"Format the content for publication. Add an SEO title (<70 chars), meta description (<160 chars), h2/h3 where useful, and a short CTA. Content: {content}"
)
PUBLISH_PROMPT = (
"Prepare publication package fields: title, summary, tags[], canonical, published_at (ISO8601 UTC), body. Content: {content}"
)
# ---------------------------
# Utilities
# ---------------------------
def utc_now_iso() -> str:
return datetime.now(timezone.utc).replace(microsecond=0).isoformat()
def json_dumps(obj: Any) -> str:
return json.dumps(obj, ensure_ascii=False, separators=(",", ":"))
def clamp_text(s: str, max_len: int = 6000) -> str:
if len(s) <= max_len:
return s
return s[: max(0, max_len - 3)] + "..."
# ---------------------------
# LLM backends
# ---------------------------
class LLM:
def complete(self, prompt: str, max_tokens: int = 800) -> str:
raise NotImplementedError
class HFInferenceLLM(LLM):
"""Hugging Face text‑generation inference. Expects env HUGGINGFACE_API_TOKEN and HUGGINGFACE_MODEL."""
def __init__(self, model: Optional[str] = None, timeout: int = 60):
self.token = os.getenv("HUGGINGFACE_API_TOKEN")
self.model = model or os.getenv("HUGGINGFACE_MODEL", "gpt2")
self.timeout = timeout
self.endpoint = f"https://api-inference.huggingface.co/models/{self.model}"
def complete(self, prompt: str, max_tokens: int = 800) -> str:
if not self.token:
raise RuntimeError("HUGGINGFACE_API_TOKEN not set")
headers = {"Authorization": f"Bearer {self.token}", "Accept": "application/json"}
payload = {"inputs": prompt, "parameters": {"max_new_tokens": max_tokens, "return_full_text": False}}
r = requests.post(self.endpoint, headers=headers, json=payload, timeout=self.timeout)
r.raise_for_status()
data = r.json()
# Response shape can vary; normalize
if isinstance(data, list) and data and "generated_text" in data[0]:
return str(data[0]["generated_text"]).strip()
if isinstance(data, dict) and "generated_text" in data:
return str(data["generated_text"]).strip()
# Fallback parsing
return json_dumps(data)
class OpenAILLM(LLM):
"""OpenAI responses via /v1/chat/completions. Requires OPENAI_API_KEY and OPENAI_MODEL."""
def __init__(self, model: Optional[str] = None, timeout: int = 60):
self.key = os.getenv("OPENAI_API_KEY")
self.model = model or os.getenv("OPENAI_MODEL", "gpt-4o-mini")
self.timeout = timeout
self.url = os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1/chat/completions")
def complete(self, prompt: str, max_tokens: int = 800) -> str:
if not self.key:
raise RuntimeError("OPENAI_API_KEY not set")
headers = {"Authorization": f"Bearer {self.key}", "Content-Type": "application/json"}
payload = {
"model": self.model,
"messages": [{"role": "user", "content": prompt}],
"temperature": 0.2,
"max_tokens": max_tokens,
}
r = requests.post(self.url, headers=headers, json=payload, timeout=self.timeout)
r.raise_for_status()
data = r.json()
return data["choices"][0]["message"]["content"].strip()
class GenericHTTPJSONLLM(LLM):
"""POSTs to LLM_ENDPOINT with JSON {prompt,max_tokens}. Expects plain text in response body or JSON {text:...}."""
def __init__(self, endpoint: Optional[str] = None, timeout: int = 60):
self.endpoint = endpoint or os.getenv("LLM_ENDPOINT")
self.timeout = timeout
def complete(self, prompt: str, max_tokens: int = 800) -> str:
if not self.endpoint:
raise RuntimeError("LLM_ENDPOINT not set")
r = requests.post(self.endpoint, json={"prompt": prompt, "max_tokens": max_tokens}, timeout=self.timeout)
r.raise_for_status()
try:
data = r.json()
return str(data.get("text") or data.get("output") or data).strip()
except Exception:
return r.text.strip()
class RuleBasedLLM(LLM):
"""Offline, deterministic fallback. Produces concise templates to keep the pipeline functional without keys."""
def complete(self, prompt: str, max_tokens: int = 800) -> str:
# Very small heuristics to keep output useful and testable
if "Generate one viral content idea" in prompt:
return "AI Side‑Hustles in 2025: 11 Practical Plays That Actually Work"
if "You are researching" in prompt:
topic = re.search(r"researching:\s*(.+?)\.\s*Summarize", prompt)
t = topic.group(1) if topic else "the topic"
return (
f"- Definition and scope of {t}\n"
f"- 2024–2025 trendline and adoption\n"
f"- 3 data points with sources\n"
f"- Risks, regulation, and future outlook"
)
if "Evaluate content quality" in prompt:
return json_dumps({
"per_criterion": {
"engagement": 8, "accuracy": 7, "originality": 7,
"emotion": 7, "readability": 8, "headline": 7,
},
"overall": 7.5,
"improvements": [
"Tighten hook with concrete stat",
"Add one contrarian insight",
"Replace generic CTA with a next‑step checklist",
],
})
if "Improve the content" in prompt:
return "[Improved] " + clamp_text(prompt.split("Content:", 1)[-1].strip())
if "Format the content for publication" in prompt:
return (
"SEO Title: Practical AI Side‑Hustles for 2025\n"
"Meta: A concise guide with data, risks, and an action checklist.\n"
"\n## Introduction\n...\n\n## CTA\nGrab the checklist."
)
if "Prepare publication package" in prompt:
now = utc_now_iso()
return json_dumps({
"title": "Practical AI Side‑Hustles for 2025",
"summary": "Concise, data‑guided ideas with risks and a checklist.",
"tags": ["AI", "side‑hustle", "2025"],
"canonical": "",
"published_at": now,
"body": "...",
})
# Default short echo
return clamp_text("[draft] " + prompt[-max_tokens:])
def build_llm() -> LLM:
# Order of preference: explicit endpoint, OpenAI, HF, fallback
try:
if os.getenv("LLM_ENDPOINT"):
logger.info("Using GenericHTTPJSONLLM")
return GenericHTTPJSONLLM()
if os.getenv("OPENAI_API_KEY"):
logger.info("Using OpenAILLM")
return OpenAILLM()
if os.getenv("HUGGINGFACE_API_TOKEN"):
logger.info("Using HFInferenceLLM")
return HFInferenceLLM()
except Exception as e:
logger.warning("LLM backend init failed, falling back: %s", e)
logger.info("Using RuleBasedLLM fallback")
return RuleBasedLLM()
# ---------------------------
# Research utilities
# ---------------------------
def ddg_search_snippets(query: str, limit: int = 5, timeout: int = 15) -> List[Dict[str, str]]:
"""Very light HTML scrape of DuckDuckGo HTML to avoid heavy APIs. Returns [{title,url,snippet}]"""
try:
url = "https://html.duckduckgo.com/html/"
r = requests.post(url, data={"q": query}, timeout=timeout, headers={"User-Agent": "agent/1.0"})
r.raise_for_status()
html = r.text
# naive parsing
results = []
for m in re.finditer(r'<a[^>]+class="result__a"[^>]*href=\"([^\"]+)\"[^>]*>(.*?)</a>', html):
link = m.group(1)
title = re.sub("<.*?>", "", m.group(2))
results.append({"title": title, "url": link, "snippet": ""})
if len(results) >= limit:
break
return results
except Exception as e:
logger.warning("ddg_search_snippets failed: %s", e)
return []
def wikipedia_summary(topic: str, timeout: int = 15) -> Optional[str]:
try:
api = "https://en.wikipedia.org/api/rest_v1/page/summary/" + requests.utils.quote(topic)
r = requests.get(api, timeout=timeout, headers={"User-Agent": "agent/1.0"})
if r.status_code == 200:
data = r.json()
return data.get("extract")
except Exception as e:
logger.warning("wikipedia_summary failed: %s", e)
return None
# ---------------------------
# Tools
# ---------------------------
@dataclass
class AgentHistory:
items: List[str] = field(default_factory=list)
def add(self, line: str) -> None:
self.items.append(line)
def render(self, max_len: int = 4000) -> str:
text = "\n".join(self.items)
return clamp_text(text, max_len)
@dataclass
class AgentConfig:
seed: int = 42
max_loops: int = 8
max_tokens: int = 800
log_jsonl: Optional[str] = os.getenv("AGENT_LOG_JSONL")
class ViralAgent:
ALLOWED_TOOLS = {
"GENERATE_IDEA",
"RESEARCH",
"GENERATE_CONTENT",
"SELF_EVALUATE",
"IMPROVE_CONTENT",
"FORMAT_CONTENT",
"PUBLISH",
"COMPLETE",
}
def __init__(self, llm: Optional[LLM] = None, cfg: Optional[AgentConfig] = None):
self.llm = llm or build_llm()
self.cfg = cfg or AgentConfig()
random.seed(self.cfg.seed)
self.history = AgentHistory()
self.session_id = uuid.uuid4().hex[:8]
logger.info("session=%s seed=%s", self.session_id, self.cfg.seed)
# -------- action loop --------
ACTION_RE = re.compile(r"^\s*action:\s*([A-Z_]+)\s*\naction_input=(.*)", re.S)
def run(self, task: str, purpose: str = "Generate viral content") -> Dict[str, Any]:
self.history.add(f"task: {task}")
context = PREFIX + f"Current Date/Time: {utc_now_iso()}\nPurpose: {purpose}\n"
for step in range(1, self.cfg.max_loops + 1):
prompt = (
f"{context}\nHistory:\n{self.history.render()}\n\n"
"Decide next step. Output exactly two lines:\n"
"action: <TOOL>\n"
"action_input=<TEXT>\n"
)
raw = self.llm.complete(prompt, max_tokens=self.cfg.max_tokens)
tool, payload = self._parse_action(raw)
logger.info("step=%s tool=%s", step, tool)
obs = self._dispatch(tool, payload, task)
self.history.add(f"observation: {clamp_text(obs, 800)}")
if tool == "COMPLETE":
return {"status": "ok", "session": self.session_id, "history": self.history.items}
return {"status": "max_loops", "session": self.session_id, "history": self.history.items}
# -------- parsing and dispatch --------
def _parse_action(self, text: str) -> Tuple[str, str]:
m = self.ACTION_RE.search(text or "")
if not m:
logger.warning("action parse failed; default to GENERATE_IDEA")
return "GENERATE_IDEA", "general tech trends 2025"
tool = m.group(1).strip().upper()
payload = m.group(2).strip()
if tool not in self.ALLOWED_TOOLS:
logger.warning("tool not allowed: %s", tool)
tool = "GENERATE_IDEA"
# guard payload
payload = clamp_text(payload, 4000)
return tool, payload
def _dispatch(self, tool: str, payload: str, task: str) -> str:
if tool == "GENERATE_IDEA":
idea = self.generate_idea(task, payload)
self.history.add(f"thought: generated idea -> {idea}")
return idea
if tool == "RESEARCH":
notes = self.research(payload or task)
self.history.add("thought: researched topic")
return notes
if tool == "GENERATE_CONTENT":
fmt = self._guess_format(payload)
notes = self._latest_research() or "key facts unavailable"
content = self.generate_content(task, fmt, notes)
self.history.add("thought: drafted content")
return content
if tool == "SELF_EVALUATE":
content = self._latest_content() or payload
return self.evaluate(content)
if tool == "IMPROVE_CONTENT":
content, feedback = self._split_two(payload)
improved = self.improve(content, feedback)
self.history.add("thought: improved content")
return improved
if tool == "FORMAT_CONTENT":
return self.format_content(payload)
if tool == "PUBLISH":
return self.publish(payload)
if tool == "COMPLETE":
return "done"
return "noop"
# -------- tool implementations --------
def generate_idea(self, topic: str, description: str) -> str:
p = IDEA_GENERATOR_PROMPT.format(topic=topic or description, history=self.history.render())
return self.llm.complete(p, max_tokens=120)
def research(self, topic: str) -> str:
topic = topic or "general topic"
bullets = []
# Try Wikipedia summary
s = wikipedia_summary(topic)
if s:
bullets.append("Wikipedia summary: " + s)
# Try DDG snippets
for r in ddg_search_snippets(topic, limit=5):
bullets.append(f"- {r['title']}{r['url']}")
# LLM consolidation
prompt = RESEARCH_PROMPT.format(topic=topic)
llm_notes = self.llm.complete(prompt, max_tokens=200)
bullets.append(llm_notes)
notes = "\n".join(bullets)
# persist short log row
self._log_jsonl({"t": utc_now_iso(), "event": "research", "topic": topic, "notes": clamp_text(notes, 2000)})
return notes
def _guess_format(self, s: str) -> str:
s = s.lower()
for key in ["blog", "book", "review", "paper", "newsletter", "social"]:
if key in s:
return {
"blog": "blog_article",
"book": "book_chapter",
"review": "review_article",
"paper": "academic_paper",
"newsletter": "newsletter",
"social": "social_media_post",
}[key]
return "blog_article"
def generate_content(self, topic: str, format_type: str, research: str) -> str:
p = CONTENT_PROMPT.format(topic=topic, format_type=format_type, research=clamp_text(research, 2000))
content = self.llm.complete(p, max_tokens=700)
self._log_jsonl({"t": utc_now_iso(), "event": "content", "format": format_type, "len": len(content)})
return content
def evaluate(self, content: str) -> str:
p = EVALUATE_PROMPT.format(content=clamp_text(content, 2500))
out = self.llm.complete(p, max_tokens=220)
# validate JSON when possible
try:
obj = json.loads(out)
if isinstance(obj, dict):
out = json_dumps(obj)
except Exception:
pass
self._log_jsonl({"t": utc_now_iso(), "event": "evaluate"})
return out
def improve(self, content: str, feedback: str) -> str:
p = IMPROVE_PROMPT.format(content=clamp_text(content, 2500), feedback=clamp_text(feedback, 800))
out = self.llm.complete(p, max_tokens=700)
self._log_jsonl({"t": utc_now_iso(), "event": "improve"})
return out
def format_content(self, content: str) -> str:
p = FORMAT_PROMPT.format(content=clamp_text(content, 2500))
out = self.llm.complete(p, max_tokens=300)
self._log_jsonl({"t": utc_now_iso(), "event": "format"})
return out
def publish(self, content: str) -> str:
p = PUBLISH_PROMPT.format(content=clamp_text(content, 2000))
out = self.llm.complete(p, max_tokens=220)
# ensure minimal JSON shape
try:
obj = json.loads(out)
if "published_at" not in obj:
obj["published_at"] = utc_now_iso()
out = json_dumps(obj)
except Exception:
# wrap as minimal manifest
out = json_dumps({"title": "Untitled", "summary": "", "tags": [], "canonical": "", "published_at": utc_now_iso(), "body": out})
self._log_jsonl({"t": utc_now_iso(), "event": "publish"})
return out
# -------- helpers --------
def _split_two(self, block: str) -> Tuple[str, str]:
parts = block.split("\n\n", 1)
if len(parts) == 2:
return parts[0].strip(), parts[1].strip()
return block, ""
def _latest_research(self) -> Optional[str]:
for line in reversed(self.history.items):
if line.startswith("observation:") and ("Wikipedia summary:" in line or line.strip().startswith("- ")):
return line.split("observation:", 1)[-1].strip()
return None
def _latest_content(self) -> Optional[str]:
for line in reversed(self.history.items):
if line.startswith("observation:") and len(line) > 30 and ("##" in line or "#" in line or "\n" in line):
return line.split("observation:", 1)[-1].strip()
return None
def _log_jsonl(self, row: Dict[str, Any]) -> None:
path = self.cfg.log_jsonl
if not path:
return
try:
with open(path, "a", encoding="utf-8") as f:
f.write(json_dumps(row) + "\n")
except Exception as e:
logger.warning("jsonl log failed: %s", e)
# ---------------------------
# CLI
# ---------------------------
def run_cli() -> None:
import argparse
parser = argparse.ArgumentParser(description="Viral content agent")
parser.add_argument("task", help="Task to execute, e.g., 'Write a blog about X'")
parser.add_argument("--purpose", default="Generate viral content")
parser.add_argument("--seed", type=int, default=int(os.getenv("AGENT_SEED", "42")))
parser.add_argument("--max-loops", type=int, default=int(os.getenv("AGENT_MAX_LOOPS", "6")))
parser.add_argument("--log-jsonl", default=os.getenv("AGENT_LOG_JSONL"))
args = parser.parse_args()
cfg = AgentConfig(seed=args.seed, max_loops=args.max_loops, log_jsonl=args.log_jsonl)
agent = ViralAgent(cfg=cfg)
result = agent.run(task=args.task, purpose=args.purpose)
print(json_dumps(result))
if __name__ == "__main__":
run_cli()