✍️ Blog Writing Agent

Turn any topic into a fully researched, citation-grounded blog post - with you in the loop at every critical step.

Multi-node agentic pipeline · Human-in-the-Loop · Groq & Ollama · Tavily web research

Quick Start  ·  Architecture  ·  Screenshots  ·  Version History  ·  Roadmap

---

What Is This?

Blog Writing Agent (BWA) is an end-to-end agentic pipeline that transforms a single blog topic into a polished, publication-ready Markdown post. It autonomously decides how to research the topic, plans a structured outline, writes each section in parallel, and assembles the final piece - all while keeping you in control at two critical Human-in-the-Loop checkpoints.

Built with LangGraph and Streamlit, BWA evolved across four progressive versions. The final version (BlogAgent-Final/) is the flagship - a production-ready web app with dual LLM provider support, intelligent routing, live web research, and a rich dark-mode UI.

---

✨ Highlights at a Glance

Feature	Details
🧭 Smart Router	Classifies your topic and picks the optimal research mode automatically
🌐 Live Web Research	Tavily-powered search with recency filtering and URL deduplication
🛑 Human-in-the-Loop	Two approval checkpoints - review strategy & outline before writing
⚡ Dual LLM Providers	Switch between Groq (cloud) or Ollama (local) in one click
🖊️ Parallel Section Writing	Workers write each blog section concurrently via LangGraph fan-out
🔍 Citation Grounding	Evidence items are sourced, deduplicated, and woven into the blog
🎨 Polished Dark UI	Streamlit app with gradient aesthetics, live logs, and blog preview
📥 One-Click Export	Download the final post as a .md file instantly

---

Screenshots

Checkpoint 1 - Review Research Strategy The Router analyses your topic, picks a research mode (open_book / hybrid / closed_book), and surfaces the planned Tavily search queries for your approval — before a single web request is made.	Generation Complete - Blog Preview After writing, the app shows word count, section count, research mode, and a styled blog preview with tabs for Raw Markdown and the full Generation Log.

---

Architecture

The pipeline follows the Router → Research → Orchestrator → Workers (fan-out) → Reducer (fan-in) agentic pattern, with two Human-in-the-Loop gates that give you full control before committing to expensive operations.

The Three Research Modes

The Router Node classifies every topic automatically - no manual selection needed:

Mode	When it applies	Recency window	Live research
closed_book	Evergreen concepts, fundamentals, theory	-	❌
hybrid	Mostly evergreen, benefits from recent examples or version names	45 days	✅
open_book	Volatile topics - news, rankings, latest releases, policy changes	7 days	✅

Node Breakdown

Node	Role	Key output
Router	Classifies topic → picks mode + generates search queries	RouterDecision
Research	Runs Tavily searches, filters & deduplicates evidence	List[EvidenceItem]
Orchestrator	Generates structured blog plan with 5–9 sections	Plan
Worker × N	Writes each section in parallel with evidence context	(task_id, section_md)
Reducer	Sorts sections, joins, runs an editor polish pass	final_blog.md

---

Quick Start

Prerequisites

• Python 3.10+
• Ollama (optional - for local inference)
• A free Groq API key (recommended for speed)
• A free Tavily API key (required for web research)

Installation

# 1. Clone the repository
git clone https://github.com/Darsh-Nandu/Blog-Writing-Agent.git
cd Blog-Writing-Agent

# 2. Install dependencies
pip install -r requirements.txt

# 3. Set up your API keys
cp .env.example .env
# Edit .env and add:
#   GROQ_API_KEY=your_groq_key_here
#   TAVILY_API_KEY=your_tavily_key_here

# 4. (Optional) Pull a local Ollama model
ollama pull llama3.1

# 5. Launch the app
cd BlogAgent-Final
streamlit run app.py

The app opens at http://localhost:8501 - enter a topic and click Generate Blog.

CLI Usage (no UI)

from nodes import run_pipeline

blog_md = run_pipeline(
    topic="State of open-source LLMs in 2026",
    provider="groq",                        # or "ollama"
    model_name="llama-3.3-70b-versatile",
)

print(blog_md)

---

Configuration

All settings live in the Streamlit sidebar - no code changes needed.

LLM Providers

Provider	Models	Best for
Groq · Cloud	Llama 3.3 70B Versatile (recommended), Llama 3.1 8B Instant, Mixtral 8×7B 32K, Gemma 2 9B	Speed, quality
Ollama · Local	Llama 3.1 8B (recommended), Llama 3.2 3B, Mistral 7B, Phi-3 Mini, Gemma 2 9B, DeepSeek R1 7B	Privacy, offline use

Environment Variables

GROQ_API_KEY=gsk_...        # Required for Groq provider
TAVILY_API_KEY=tvly-...     # Required for web research (hybrid / open_book modes)

---

Human-in-the-Loop Checkpoints

BWA 4.0 pauses at two checkpoints before committing to expensive operations.

🛑 Checkpoint 1 - Review Research Strategy

Triggered after routing. Shows:

• Mode chosen (Open Book / Hybrid / Closed Book)
• Research needed (Yes / No) and recency window (7 / 45 days)
• Planned search queries - exactly what will be sent to Tavily

Add optional feedback ("add a query about LangGraph v0.4") before approving or going back to the topic.

🛑 Checkpoint 2 - Review Blog Outline

Triggered after the orchestrator generates the plan. Shows:

• Blog title, audience, tone, and estimated word count
• Every section card with goal, sub-bullets, word target, and flags (code / citations / research)
• Collected evidence items with source URLs and dates

Approve, Regenerate the plan, or Go back to tweak the research strategy.

---

Version History

Feature	BWA 1.0	BWA 2.0	BWA 3.0	BWA 4.0
Core Orchestrator → Workers → Reducer	✅	✅	✅	✅
Parallel section writing	✅	✅	✅	✅
Structured section planning	❌	✅	✅	✅
Technical prompts (goals, bullets, word counts)	❌	✅	✅	✅
Routing (open_book / hybrid / closed_book)	❌	❌	✅	✅
Live web research (Tavily)	❌	❌	✅	✅
Citation grounding	❌	❌	✅	✅
Recency filtering	❌	❌	✅	✅
Externalized prompts	❌	❌	✅	✅
Human-in-the-Loop checkpoints	❌	❌	❌	✅
Streamlit web UI	❌	❌	❌	✅
Dual LLM provider (Groq + Ollama)	❌	❌	❌	✅
Blog preview + markdown download	❌	❌	❌	✅
Editor polish pass (reducer)	❌	❌	❌	✅

BWA 1.0 establishes the core Orchestrator → Fan-Out → Fan-In pattern with LangGraph. No research, no structured planning — but the essential pipeline is there.

BWA 2.0 introduces structured section planning via Pydantic schemas. Each section gets a goal, word count, and sub-bullets. Workers now know what to write, not just how much.

BWA 3.0 adds the Router node, Tavily integration, recency filtering, citation grounding, and externalized prompt modules in prompts.py for easy tuning.

BWA 4.0 wraps the entire pipeline in a Streamlit UI with a polished dark-mode design, dual LLM provider support, two HITL checkpoints, live generation logs, blog preview tabs, and one-click .md download.

---

Sample Output

BWA 4.0 / 3.0 - "State of Open Source LLMs in 2026" (open_book, research-grounded)

Market Landscape & Leading Models

Open-source LLMs have reached near-parity with closed frontier models on a range of benchmarks. Meta's Llama 3 family continues to dominate downloads, while Mistral AI maintains strong community adoption.

Deployment & Inference Trends

The introduction of AI-native platforms has democratised access - quantised models running on consumer hardware are now a practical production path. vLLM and llama.cpp have become the de-facto serving stacks for self-hosted inference.

All source URLs are real, fetched, and deduplicated. Workers cannot invent citations.

BWA 2.0 - "Mastering Self-Attention in Transformers" (closed_book, technical deep-dive)

The Self-Attention Mechanism

Self-attention computes relationships between every pair of tokens: α = softmax(QK^T / √d_k) · V

class ScaledDotProductAttention(nn.Module):
    def forward(self, Q, K, V):
        scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.d_k)
        weights = F.softmax(scores, dim=-1)
        return torch.matmul(weights, V)

---

Repository Structure

Blog-Writing-Agent/
│
├── BWA_1.0/                    # v1 · Core fan-out/fan-in pipeline
│   ├── app.py
│   ├── nodes.py
│   ├── custom_objects.py
│   ├── requirements.txt
│   └── README.md
│
├── BWA_2.0/                    # v2 · Richer prompts, structured planning
│   ├── app.py
│   ├── nodes.py
│   ├── custom_objects.py
│   ├── BWA_2.0_test_1.md       ← sample output
│   ├── requirements.txt
│   └── README.md
│
├── BWA_3.0/                    # v3 · Router, Tavily research, citations
│   ├── app.py
│   ├── nodes.py
│   ├── custom_objects.py
│   ├── prompts.py
│   ├── requirements.txt
│   └── README.md
│
├── BlogAgent-Final/            # v4 · ⭐ FLAGSHIP — Streamlit UI + HITL
│   ├── app.py                  ← Streamlit app (UI + session flow)
│   ├── nodes.py                ← All pipeline nodes
│   ├── llm_factory.py          ← Groq + Ollama provider abstraction
│   ├── custom_objects.py       ← Pydantic schemas (State, Plan, Task, Evidence…)
│   └── prompts.py              ← All LLM system prompts
│
├── assets/                     ← Screenshots & diagrams
│   ├── screenshot_checkpoint.png
│   ├── screenshot_output.png
│   └── architecture.svg
│
├── requirements.txt
└── README.md                   ← This file

---

Tech Stack

	Tool	Version	Purpose
🔗	LangGraph	≥ 0.2	Agent graph, parallel fan-out, Send() API
🦜	LangChain	≥ 0.3	LLM interface, structured output
⚡	Groq	≥ 0.11	Cloud LLM inference (ultra-fast)
🦙	Ollama	any	Local LLM inference - fully offline
🔍	Tavily	≥ 0.3	Live web research & evidence retrieval
🛡️	Pydantic v2	≥ 2.0	Typed schemas for all pipeline objects
🎨	Streamlit	≥ 1.35	Web UI with dark-mode design system
🐍	Python	3.10+	Runtime

---

Roadmap

✅ BWA 1.0  - Core fan-out / fan-in pipeline
✅ BWA 2.0  - Structured prompts & section planning
✅ BWA 3.0  - Web research, routing, citation grounding
✅ BWA 4.0  - Human-in-the-Loop · Streamlit UI · Groq + Ollama

⬜ BWA 5.0  - Output quality evaluation node (auto-score)
⬜ BWA 5.0  - Multi-format export  (HTML · PDF · Notion · Medium)
⬜ BWA 5.0  - Persistent history & blog library
⬜ BWA 5.0  - Image generation per section (DALL-E / SDXL)
⬜ BWA 5.0  - Custom tone / style profiles
⬜ BWA 5.0  - Docker + one-command deploy

---

Contributing

Contributions are welcome! If you find a bug, have a feature idea, or want to improve the prompts:

1. Fork the repository
2. Create a feature branch: git checkout -b feat/my-feature
3. Commit your changes: git commit -m "feat: add my feature"
4. Push and open a Pull Request

---

MIT License  ·  Built with ❤️ by Darsh Nandu

Star ⭐ the repo if you find it useful!