# LangGraph × Multi-LLM Book Builder

Generate a book from a short problem statement using a graph of LLM steps:
**spec → toc → plan → draft → images → assemble**.

---

## Why this exists

Large language models often return malformed or chatty JSON. This project was designed to be **fault-tolerant**:

- Extracts JSON fences, but also parses the **first JSON value only**.
- Repairs malformed outputs with a conversion prompt.
- Normalizes output before optional schema validation.
- Never hard-fails: TOC has heuristic parsing and local synthesis fallback.
- Checkpoints runs (`spec.json`, `toc.json`, etc.) for `--resume`.

---

## Features

- **Multi-LLM routing**
  - Ollama: `spec`, `toc`, `plan`, `images`
  - Gemini Flash: `draft`, `assemble`
- **Robust JSON handling**
  - Extract fenced JSON, parse only first JSON, repair when needed
- **Normalization**
  - `spec`: coerce audience/goals/constraints
  - `plan`: coerce objectives/key_ideas/image_prompts
- **TOC resilience**
  - Unwrap wrappers (`{"toc": [...]}`)
  - Parse numbered lines
  - Synthesize fallback TOC with rebalanced pages
- **CLI Options**
  - `--resume`, `--dry_run`, `--sample_chapters`, `--max_workers`, `--strict_json`

---

## Setup

```bash
python -m venv .venv && source .venv/bin/activate
pip install -U langgraph langchain-core langchain-ollama langchain-google-genai tenacity jsonschema
ollama pull llama3.1:8b
export GOOGLE_API_KEY="YOUR_KEY"

---

Usage

Minimal run:

python book.py \
  --problem "How to start a bootstrapped SaaS" \
  --pages 200 \
  --model llama3.1:8b

Flags:

• --out ./book_out output dir (default ./book_out)
• --resume reuse existing outputs
• --dry_run skip draft/assemble
• --sample_chapters 2 draft only N chapters (0 = all)
• --max_workers 4 concurrent drafting threads
• --base_url http://127.0.0.1:11434 custom Ollama URL
• --google_api_key ... override env var
• --strict_json enforce schema strictly

Outputs:

spec.json
toc.json
plans.json
drafts.json
image_prompts.json
book.md

---

Architecture

LangGraph StateGraph
  spec  ->  toc  ->  plan  ->  draft  ->  images  ->  assemble  -> END

• Structured nodes → Ollama
• Heavy drafting/assembly → Gemini Flash

---

Challenges and Fixes

1. Schema mismatches (goals_outcomes vs goals)
   • Fix: Normalize first, validate later.
2. Non-JSON outputs (markdown, prose)
   • Fix: Extract fences, parse first JSON, repair with conversion prompt.
3. JSONDecodeError: Extra data
   • Fix: Parse only first JSON value, ignore trailing text.
4. Ollama client temperature error
   • Fix: Avoid passing temperature param.
5. TOC hard failures
   • Fix: unwrap → heuristic numbered lines → fallback synthesis.
6. Resume loading bad files
   • Fix: Normalize on load; delete broken files to regenerate.

---

Potential Pitfalls

• Wrong Ollama URL or missing model → pull model with ollama pull.
• Missing/invalid GOOGLE_API_KEY → set env var or pass CLI.
• Version mismatches between langchain-ollama and ollama client.
• Very long inputs exceeding context → trim pages/words.
• Too high concurrency → API rate limits.

---

Troubleshooting

• Crash at TOC → delete toc.json, rerun.
• Extra data error → ensure first-JSON parsing patch is applied.
• Ollama temperature error → update client, don’t pass temperature.
• Gemini auth error → check GOOGLE_API_KEY, confirm gemini-2.5-flash model.

---

Extending

• Export per-chapter markdown
• Add image generation
• Add revise/review nodes
• Integrate retrieval for references/examples

---

License

MIT

---

Acknowledgements

• LangGraph
• LangChain Ollama
• LangChain Google Generative AI