AutoGenBook (CLI)

AutoGenBook is a modular Python CLI for generating long-form documents from short specifications using mode-specific pipelines (book, paper, presentation, scientist, proposal, reviewer). Each pipeline builds a document graph, injects retrieval context, writes leaf sections with agent prompts, and assembles Markdown by default (book/paper) with optional LaTeX/PDF conversion; legacy LaTeX-first generation remains available. (main.py:parse_args, autogenbook/orchestrator.py:run, autogenbook/pipelines/*, book_builder.py:build_markdown_document, book_builder.py:build_latex_document, book_builder.py:compile_pdf)

This project exists to make long-form generation repeatable and auditable: structure is explicit (graph nodes with page budgets), sources are tracked via stable IDs, and optional audits can flag missing citations or unsupported numeric claims. (book_builder.py:build_graph_from_book_json, autogenbook/graph/doc_graph.py, rag_kb.py:KnowledgeBase, autogenbook/retrieval/types.py:RetrievalItem, autogenbook/audit/latex_auditor.py:audit_latex)

Who it is for

• Authors drafting structured books or reports with reproducible section boundaries. (book_builder.py:generate_contents)
• Researchers drafting papers with explicit citations and related-work scaffolding. (autogenbook/pipelines/paper_pipeline.py:run_paper)
• Speakers drafting slide decks with grounded, concise talking points. (autogenbook/pipelines/presentation_pipeline.py:run_presentation)
• Teams running an experiment + writing loop from the same run artifacts. (autogenbook/pipelines/scientist_pipeline.py:run_scientist)
• Grant proposal and review workflows that need MCP-tool citations and traceability. (autogenbook/pipelines/proposal_pipeline.py:run_proposal, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)

Modes at a glance

Mode	Primary input	Core steps	Key outputs
book	TXT spec (--input) or JSON (--json)	TXT -> JSON -> graph -> sections -> Markdown (default) -> optional LaTeX/PDF (pandoc)	structure_graph.json, sections/*.md, <title>.md, optional .tex/.pdf
paper	TXT spec (--input) or JSON (--json)	TXT -> JSON -> graph -> sections -> Markdown (default) -> LaTeX/PDF (pandoc)	structure_graph.json, sections/*.md, related_work.json, <title>.md, optional .tex/.pdf, refs.bib (BibTeX)
presentation	TXT spec (--input) or JSON (--json)	TXT -> JSON -> graph -> slides -> Markdown (default) -> optional PPTX / Beamer LaTeX/PDF + narration/audio/video	structure_graph.json, slides/*.md, images/*.png, <title>.md, optional .pptx / .tex / .pdf, narration .md/.json, audio .wav, video .mp4
scientist	Template + optional KB	Idea -> literature -> plan -> experiment -> analysis -> write -> review -> revise	experiments/<run_id>/metrics.json, review.json, audit_report.json, LaTeX outputs
proposal	proposal_input.txt + KB1/KB2	Requirements -> outline -> sections w/ MCP citations -> final review	outline/, draft/, final/, per-section artifacts, proposal_final.md
reviewer	KB2 (+ optional KB1)	Compose reviewer prompt -> draft review	review_final.md (+ optional PDF/TeX via pandoc)

How it works (core flow)

1. Parse CLI args and create a run context (output paths + run id). (main.py:parse_args, autogenbook/state.py:RunContext)
2. Load the prompt pack for the selected mode. (autogenbook/prompts/*_loader.py, autogenbook/prompts/registry.py:set_prompt_registry)
3. Build or load a structure graph (book/paper/presentation). (book_builder.py:generate_book_json_from_txt, book_builder.py:build_graph_from_book_json, autogenbook/pipelines/paper_pipeline.py:_build_graph_from_paper_json, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
4. Subdivide oversized nodes to leaf sections. (book_builder.py:subdivide_graph, autogenbook/pipelines/paper_pipeline.py:_subdivide_paper_graph)
5. Generate leaf section content with retrieval context and agent prompts. (book_builder.py:generate_contents, autogenbook/agents/base.py:BaseAgent, autogenbook/retrieval/manager.py:RetrievalManager)
6. Assemble Markdown (book/paper default) or LaTeX (legacy), then optionally convert via pandoc and compile PDF. (book_builder.py:build_markdown_document, book_builder.py:build_latex_document, book_builder.py:compile_pdf)

Key features

• Document graph planning with explicit section budgets. (book_builder.py:build_graph_from_book_json, autogenbook/graph/doc_graph.py)
• Local RAG knowledge base built from PDF/DOCX/PPTX/MD/TXT with BM25. (rag_kb.py:KnowledgeBase, rag_kb.py:SUPPORTED_EXTS)
• Optional web retrieval via MCP paper tools and/or Tavily search. (autogenbook/retrieval/mcp_papers.py:MCPPaperRetriever, autogenbook/retrieval/tavily.py:TavilyRetriever)
• Agent outputs validated against Pydantic schemas with optional repair. (autogenbook/agents/base.py:BaseAgent, autogenbook/schemas/*)
• LaTeX auditing for unknown citations, missing figures, and numeric evidence gaps. (autogenbook/audit/latex_auditor.py:audit_latex)
• Resumable runs via structure_graph.json and per-section files. (autogenbook/graph/doc_graph.py:save_graph_json, book_builder.py:generate_contents)

Requirements

• Python dependencies are declared in requirements.txt. (requirements.txt)
• OpenRouter API key is required when using OpenRouter; OpenAI-compatible endpoints can be set via --llm-base-url or AUTOGENBOOK_LLM_BASE_URL. (openrouter_llm.py:OpenRouterLLM.__init__, main.py:parse_args)
• PDF output requires LuaLaTeX (lualatex). (book_builder.py:compile_pdf)
• Book/paper Markdown→TeX/PDF conversion uses pandoc. Proposal/reviewer conversions also use pandoc when available. (book_builder.py:_convert_markdown_to_latex, autogenbook/pipelines/proposal_pipeline.py:_convert_with_pandoc, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
• The scientist template uses scikit-learn and matplotlib; fallback PNG is written if matplotlib is unavailable. (autogenbook/templates/toy_classification/run_experiment.py:_run_sklearn, _plot)

Install

pip install -r requirements.txt

Source: requirements.txt

Quickstart

Set the OpenRouter key:

export OPENROUTER_API_KEY="YOUR_KEY"

Source: openrouter_llm.py:OpenRouterLLM.__init__

Use a local OpenAI-compatible server (e.g., LM Studio):

export AUTOGENBOOK_LLM_BASE_URL="http://localhost:1234/v1"
# Optional if your server requires a key:
export AUTOGENBOOK_LLM_API_KEY="local-key"

Source: openrouter_llm.py:OpenRouterLLM.__init__

Book mode (sample input; Markdown-first):

python main.py --mode book --input input/book/book_input.txt --out-dir output/book/out_book

Source: main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book

Book mode now emits Markdown by default; add --export-tex to produce TeX/PDF via pandoc. (autogenbook/pipelines/book_pipeline.py:run_book)

Paper mode (sample input; Markdown-first):

python main.py --mode paper --input input/paper/paper_input.txt --out-dir output/paper/out_paper

Source: main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper

Paper mode now emits Markdown by default and converts to TeX/PDF via pandoc unless disabled. Use --legacy-tex to force the old LaTeX-first pipeline. (autogenbook/pipelines/paper_pipeline.py:run_paper)

Disable PDF if LuaLaTeX is unavailable:

python main.py --mode book --input input/book/book_input.txt --out-dir output/book/out_book --no-pdf

Source: book_builder.py:compile_pdf, main.py:parse_args

More examples: docs/EXAMPLES.md. (docs/EXAMPLES.md)

CLI reference (complete)

All flags are defined in main.py:parse_args. (main.py:parse_args)

Core

• --mode {book,paper,presentation,scientist,proposal,reviewer}: Select pipeline. (main.py:parse_args, autogenbook/orchestrator.py:run)
• --out-dir, -o: Output directory. (main.py:parse_args, autogenbook/state.py:RunContext)
• --llm-base-url: Override OpenAI-compatible base URL for all LLM calls in the run. (main.py:parse_args, openrouter_llm.py:OpenRouterLLM.__init__)

Inputs and structure

• --input, -i: TXT spec (book/paper/presentation). (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book, autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
• --presentation-input: Presentation input file. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
• --proposal-input: Proposal input file. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• --json, -j: Structure JSON (book/paper/presentation). (main.py:parse_args, main.py:main)
• --use-json: Use JSON without prompting. (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book, autogenbook/pipelines/paper_pipeline.py:run_paper)
• --use-txt: Regenerate JSON from TXT. (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book, autogenbook/pipelines/paper_pipeline.py:run_paper)
• --resume: Resume from previous outputs. (main.py:parse_args, book_builder.py:generate_contents, autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/proposal_pipeline.py:run_proposal)

Retrieval / KB

• --kb-dir: Local KB for book/paper/scientist. (main.py:parse_args, rag_kb.py:KnowledgeBase.build_from_directory)
• --kb1-dir, --kb2-dir: Proposal/reviewer KBs. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
• --rebuild-kb: Force KB rebuild. (main.py:parse_args, rag_kb.py:KnowledgeBase.build_from_directory)
• --enable-web-rag: Enable web retrieval (proposal requires MCP tools). (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• --web-rag-k: Number of web results. (main.py:parse_args, autogenbook/retrieval/manager.py:RetrievalManager)

Paper

• --paper-venue: Target venue string. (main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:_build_paper_json)
• --citation-style {bibtex,footnote}: Paper citation mode. (main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper)

Proposal

• --max-iters: Max iterations (proposal outline/review loop; scientist iterations). (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal, autogenbook/pipelines/scientist_pipeline.py:run_scientist)
• --section-retries: Retry attempts per proposal section. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• --min-section-citations: Minimum citations per section. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• --dont-ask / --dont_ask: Disable user prompts. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• --proposal-llm1-model .. --proposal-llm5-model: Per-role model overrides. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:_resolve_model_override)
• --proposal-llm1-base-url .. --proposal-llm5-base-url: Per-role base URL overrides. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)

Reviewer

• --reviewer-llm1-model, --reviewer-llm2-model: Model overrides. (main.py:parse_args, autogenbook/pipelines/reviewer_pipeline.py:_resolve_model_override)
• --reviewer-llm1-base-url, --reviewer-llm2-base-url: Per-role base URL overrides. (main.py:parse_args, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
• --reviewer-direct-pdf: Include direct PDF text from KB2 in prompt. (main.py:parse_args, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)

Output toggles

• --no-tex: Suppress .tex as a final artifact for book/paper/scientist (LaTeX still runs when PDF is requested; paper still generates intermediate .tex for PDF). Disable pandoc TeX conversion in reviewer. (main.py:parse_args, book_builder.py:AppConfig, autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
• --no-pdf: Disable PDF output. (main.py:parse_args, book_builder.py:AppConfig, autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/scientist_pipeline.py:run_scientist, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
• --no-md: Disable Markdown export (book/paper/scientist). (main.py:parse_args, book_builder.py:AppConfig, autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/scientist_pipeline.py:run_scientist)

Auditing and schema validation

• --audit: Enable audit (paper/scientist default on; proposal default off). (main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/scientist_pipeline.py:run_scientist, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• --audit-mode {off,warn,strict}: Audit severity. (main.py:parse_args, autogenbook/audit/latex_auditor.py:audit_latex)
• --audit-window-chars: Evidence window size for numeric claims. (main.py:parse_args, autogenbook/audit/latex_auditor.py:audit_latex)
• Strict audit stops the run before PDF emission if errors are found. (autogenbook/pipelines/paper_pipeline.py:run_paper, autogenbook/pipelines/book_pipeline.py:run_book)
• --audit-book: Enable audit in book mode. (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book)
• --audit-book-mode {off,warn,strict}: Book audit severity. (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book)
• --fail-fast-schema: Fail immediately on schema validation errors. (main.py:parse_args, autogenbook/agents/base.py:BaseAgent._validate_with_repair)

Complete reference: docs/API_REFERENCE.md. (docs/API_REFERENCE.md)

Mode-specific arguments and examples

This section groups flags by mode and shows practical command lines. All flags are defined in main.py:parse_args. (main.py:parse_args)

Sample specs in this repo live under input/ (e.g., input/book/book_input.txt, input/paper/paper_input.txt, input/presentation/presentation_input.txt, input/proposal/proposal_input.txt), so most runs pass explicit --input or --proposal-input. (input/book/book_input.txt, input/paper/paper_input.txt, input/presentation/presentation_input.txt, input/proposal/proposal_input.txt)

Book mode (--mode book)

Arguments and defaults:

Flag	Required?	Default	Notes
--input, -i	Yes (unless using --use-json)	book_input.txt	TXT spec used to build the book JSON. (main.py:parse_args, book_builder.py:generate_book_json_from_txt)
--json, -j	Optional	book_structure.json	Structure JSON; used when --use-json is set. (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book)
--use-json / --use-txt	Optional	n/a	Mutually exclusive; force JSON reuse or regeneration. (main.py:parse_args)
--kb-dir	Optional	none	Build local KB from PDF/DOCX/PPTX/MD/TXT. (rag_kb.py:KnowledgeBase.build_from_directory)
--rebuild-kb	Optional	false	Force KB rebuild when --kb-dir is set. (rag_kb.py:KnowledgeBase.build_from_directory)
--enable-web-rag, --web-rag-k	Optional	false / 5	Enable web retrieval (MCP/Tavily) and set result count; warns and falls back to KB if tools are unavailable. (autogenbook/pipelines/book_pipeline.py:run_book)
--resume	Optional	false	Skip already-generated sections (sections/*.md by default; legacy LaTeX uses .tex). (book_builder.py:generate_contents)
--export-tex	Optional	false	Book mode: export TeX/PDF from Markdown output (pandoc). (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book)
--legacy-tex	Optional	false	Book mode: use legacy LLM LaTeX generation instead of Markdown-first. (main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book)
--audit-book, --audit-book-mode	Optional	false / warn	Enable book audit and set severity. (autogenbook/pipelines/book_pipeline.py:run_book)
--no-tex, --no-pdf, --no-md	Optional	false	Disable LaTeX/PDF/Markdown outputs. (book_builder.py:AppConfig)
--out-dir, -o	Optional	./out	Output directory root for all artifacts. (main.py:parse_args, autogenbook/state.py:RunContext)

Book mode now generates Markdown sections by default and only exports TeX/PDF when --export-tex is set (requires pandoc). Use --legacy-tex to force the older LaTeX-first pipeline if conversion fails. (autogenbook/pipelines/book_pipeline.py:run_book)

Example: book with KB rebuild, web RAG, and resume:

python main.py --mode book --input input/book/book_input.txt --kb-dir input/book/kb --rebuild-kb --enable-web-rag --web-rag-k 6 --resume --out-dir output/book/out_book

Source: main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book

Example: use a prebuilt structure JSON (skip TXT conversion):

python main.py --mode book --json book_structure.json --use-json --out-dir output/book/out_json

Source: main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book

Example: force TXT regeneration + strict audit without PDF:

python main.py --mode book --input input/book/book_input.txt --use-txt --audit-book --audit-book-mode strict --no-pdf --out-dir output/book/out_rebuild

Source: main.py:parse_args, autogenbook/pipelines/book_pipeline.py:run_book

Paper mode (--mode paper)

Arguments and defaults:

Flag	Required?	Default	Notes
--input, -i	Yes (unless using --use-json)	book_input.txt	TXT spec used to build the paper JSON. (main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:_build_paper_json)
--json, -j	Optional	paper_structure.json	Default JSON switches to paper_structure.json in paper mode. (main.py:parse_args, main.py:main)
--use-json / --use-txt	Optional	n/a	Mutually exclusive; force JSON reuse or regeneration. (main.py:parse_args)
--paper-venue	Optional	arXiv	Venue label used in prompts. (main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:_build_paper_json)
--citation-style	Optional	bibtex	bibtex or footnote. (main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper)
--audit, --audit-mode, --audit-window-chars	Optional	on / warn / 600	Audit defaults to enabled; --audit-mode off disables it. (autogenbook/pipelines/paper_pipeline.py:run_paper)
--kb-dir, --rebuild-kb	Optional	none	Build local KB for retrieval. (rag_kb.py:KnowledgeBase.build_from_directory)
--enable-web-rag, --web-rag-k	Optional	false / 5	Enable web retrieval (MCP/Tavily); warns and falls back to KB if tools are unavailable. (autogenbook/pipelines/paper_pipeline.py:run_paper)
--legacy-tex	Optional	false	Use legacy LaTeX-first generation instead of Markdown-first. (autogenbook/pipelines/paper_pipeline.py:run_paper)
--no-pdf, --no-md	Optional	false	Disable PDF/Markdown outputs. (autogenbook/pipelines/paper_pipeline.py:run_paper)
--out-dir, -o	Optional	./out	Output directory root for all artifacts. (main.py:parse_args, autogenbook/state.py:RunContext)

Example: paper with custom venue, footnote citations, and strict audit:

python main.py --mode paper --input input/paper/paper_input.txt --paper-venue "ACL 2025" --citation-style footnote --audit --audit-mode strict --out-dir output/paper/out_paper

Source: main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper

Example: paper with web RAG enabled:

python main.py --mode paper --input input/paper/paper_input.txt --enable-web-rag --web-rag-k 8 --out-dir output/paper/out_web

Source: main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper

Example: reuse an existing paper structure JSON:

python main.py --mode paper --json paper_structure.json --use-json --out-dir output/paper/out_json

Source: main.py:parse_args, autogenbook/pipelines/paper_pipeline.py:run_paper

Presentation mode (--mode presentation)

Arguments and defaults:

Flag	Required?	Default	Notes
--input, -i	Yes (unless using --use-json)	presentation_input.txt	TXT spec used to build the presentation JSON. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--json, -j	Optional	presentation_structure.json	Default JSON switches to presentation_structure.json in presentation mode. (main.py:parse_args, main.py:main)
--presentation-tex	Optional	false	Convert Markdown to Beamer .tex/.pdf. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--presentation-pptx	Optional	false	Export Markdown deck to PowerPoint .pptx. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--presentation-narration	Optional	false	Generate per-slide narration text. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--presentation-narration-model	Optional	none	Model override for narration. (main.py:parse_args)
--presentation-tts	Optional	false	Generate audio from narration. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--presentation-tts-mode	Optional	openrouter	openrouter or local TTS backend. (main.py:parse_args)
--presentation-tts-model	Optional	none	OpenRouter TTS model override. Default: openai/gpt-4o-mini-tts-2025-12-15. (main.py:parse_args)
--presentation-video	Optional	false	Render video from PDF slides + audio. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--presentation-exclude-slides	Optional	empty	Exclude slide indices or ranges from audio/video (e.g., 2,5,10-12). (main.py:parse_args)
--presentation-citations	Optional	false	Include Harvard-style citations from MCP paper tools in slide text. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--presentation-image-model	Optional	none	OpenRouter image model override for slide images. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--no-image	Optional	false	Disable slide image generation entirely. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--disable-general-knowledge-citation	Optional	false	Forbid/remove General background knowledge marker from slide text. (main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--kb-dir, --rebuild-kb	Optional	none	Build local KB for retrieval. (rag_kb.py:KnowledgeBase.build_from_directory)
--enable-web-rag, --web-rag-k	Optional	false / 5	Enable web retrieval (MCP/Tavily); warns and falls back to KB if tools are unavailable. (autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--resume	Optional	false	Skip already-generated slides. (autogenbook/pipelines/presentation_pipeline.py:run_presentation)
--out-dir, -o	Optional	./out	Output directory root for all artifacts. (main.py:parse_args, autogenbook/state.py:RunContext)

Example: basic presentation with web RAG:

python main.py --mode presentation --input input/presentation/presentation_input.txt --enable-web-rag --web-rag-k 6 --out-dir output/presentation/out_presentation --disable-general-knowledge-citation --no-image

Source: main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation

Example: Beamer export + narration + audio:

python main.py --mode presentation --input input/presentation/presentation_input.txt --presentation-pptx --presentation-tex --presentation-narration --presentation-tts --out-dir output/presentation/out_media

Source: main.py:parse_args, autogenbook/pipelines/presentation_pipeline.py:run_presentation

Note: Beamer export requires pandoc + pypandoc. Audio/video features require optional packages such as pymupdf, moviepy, pydub, soundfile, scipy, TTS, and ffmpeg. Slide image generation uses an OpenRouter image model (override with --presentation-image-model, disable with --no-image).

Scientist mode (--mode scientist)

Arguments and defaults:

Flag	Required?	Default	Notes
--max-iters	Optional	1	Number of patch iterations. (autogenbook/pipelines/scientist_pipeline.py:DEFAULT_MAX_ITERS, autogenbook/pipelines/scientist_pipeline.py:run_scientist)
--kb-dir, --rebuild-kb	Optional	none	Build local KB for retrieval. (rag_kb.py:KnowledgeBase.build_from_directory)
--enable-web-rag, --web-rag-k	Optional	false / 5	Enable web retrieval (MCP/Tavily); warns and falls back to KB if tools are unavailable. (autogenbook/pipelines/scientist_pipeline.py:run_scientist)
--audit, --audit-mode, --audit-window-chars	Optional	on / strict / 600	Audit defaults to enabled; scientist forces strict when --audit-mode is omitted. (autogenbook/pipelines/scientist_pipeline.py:run_scientist)
--no-pdf, --no-md	Optional	false	Disable PDF/Markdown outputs. (autogenbook/pipelines/scientist_pipeline.py:run_scientist)
--out-dir, -o	Optional	./out	Output directory root for all artifacts. (main.py:parse_args, autogenbook/state.py:RunContext)

Example: run two scientist iterations without PDF:

python main.py --mode scientist --max-iters 2 --no-pdf --out-dir output/scientist/out_scientist

Source: main.py:parse_args, autogenbook/pipelines/scientist_pipeline.py:run_scientist

Example: disable audit and enable web retrieval:

python main.py --mode scientist --audit-mode off --enable-web-rag --web-rag-k 5 --out-dir output/scientist/out_web

Source: main.py:parse_args, autogenbook/pipelines/scientist_pipeline.py:run_scientist

Proposal mode (--mode proposal)

Arguments and defaults:

Flag	Required?	Default	Notes
--proposal-input	Yes	proposal_input.txt	Proposal spec file. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--kb1-dir, --kb2-dir	Yes	none	Required KB directories. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--enable-web-rag	Yes	false	Proposal mode errors without MCP paper tools. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--web-rag-k	Optional	5	Web retrieval result count. (autogenbook/retrieval/manager.py:RetrievalManager)
--rebuild-kb	Optional	false	Force rebuild of KB1/KB2 caches. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--max-iters	Optional	3	Outline/review loop count (default set in main.py). (main.py:main)
--section-retries	Optional	3	Retry attempts per section. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--min-section-citations	Optional	1	Minimum citations per section. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--dont-ask / --dont_ask	Optional	false	Disable user prompts; rely on KB/web. (main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--proposal-llm1-model .. --proposal-llm5-model	Optional	none	Per-role model overrides. (autogenbook/pipelines/proposal_pipeline.py:_resolve_model_override)
--proposal-llm1-base-url .. --proposal-llm5-base-url	Optional	none	Per-role base URL overrides (OpenAI-compatible endpoints). (main.py:parse_args, openrouter_llm.py:OpenRouterLLM.__init__)
--audit, --audit-mode, --audit-window-chars	Optional	off / warn / 600	Final audit controls. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)
--out-dir, -o	Optional	./out	Output directory root for all artifacts. (main.py:parse_args, autogenbook/state.py:RunContext)

Example: proposal with required KBs and MCP citations:

python main.py --mode proposal --proposal-input input/proposal/proposal_input.txt --kb1-dir input/proposal/kb1 --kb2-dir input/proposal/kb2 --enable-web-rag --out-dir output/proposal/out_proposal

Source: main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal

Example: proposal in non-interactive mode with stricter citation requirements:

python main.py --mode proposal --proposal-input input/proposal/proposal_input.txt --kb1-dir input/proposal/kb1 --kb2-dir input/proposal/kb2 --enable-web-rag --min-section-citations 2 --dont-ask --out-dir output/proposal/out_strict

Source: main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal

Example: proposal with custom model overrides and loop tuning:

python main.py --mode proposal --proposal-input input/proposal/proposal_input.txt --kb1-dir input/proposal/kb1 --kb2-dir input/proposal/kb2 --enable-web-rag --max-iters 4 --section-retries 2 --proposal-llm1-model openai/gpt-5-mini --proposal-llm4-model openai/gpt-5-mini --out-dir output/proposal/out_custom

Source: main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal, openrouter_llm.py:LLMConfig

Example: proposal with per-role endpoints (LLM1/LLM4 on local server):

python main.py --mode proposal --proposal-input input/proposal/proposal_input.txt --kb1-dir input/proposal/kb1 --kb2-dir input/proposal/kb2 --enable-web-rag --proposal-llm1-base-url http://localhost:1234/v1 --proposal-llm4-base-url http://localhost:1234/v1 --out-dir output/proposal/out_local_roles

Source: main.py:parse_args, autogenbook/pipelines/proposal_pipeline.py:run_proposal, openrouter_llm.py:OpenRouterLLM.__init__

Reviewer mode (--mode reviewer)

Arguments and defaults:

Flag	Required?	Default	Notes
--kb2-dir	Yes	none	Required KB2 directory (thesis/work). (autogenbook/pipelines/reviewer_pipeline.py:build_or_load_kb2)
--kb1-dir	Optional	none	Optional KB1 directory (norms/requirements). (autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
--rebuild-kb	Optional	false	Force rebuild of reviewer KB caches. (autogenbook/pipelines/reviewer_pipeline.py:_configure_reviewer_kb)
--reviewer-direct-pdf	Optional	false	Include direct PDF text in the review prompt. (autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
--reviewer-llm1-model, --reviewer-llm2-model	Optional	none	Model overrides. (autogenbook/pipelines/reviewer_pipeline.py:_resolve_model_override)
--reviewer-llm1-base-url, --reviewer-llm2-base-url	Optional	none	Per-role base URL overrides (OpenAI-compatible endpoints). (main.py:parse_args, openrouter_llm.py:OpenRouterLLM.__init__)
--no-pdf, --no-tex	Optional	false	Skip pandoc conversions. (autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
--out-dir, -o	Optional	./out	Output directory root for all artifacts. (main.py:parse_args, autogenbook/state.py:RunContext)

Example: reviewer with direct PDF context and no conversions:

python main.py --mode reviewer --kb2-dir input/reviewer/kb2 --reviewer-direct-pdf --no-pdf --no-tex --out-dir output/reviewer/out_reviewer

Source: main.py:parse_args, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer

Example: reviewer with KB1 norms and model overrides:

python main.py --mode reviewer --kb2-dir input/reviewer/kb2 --kb1-dir input/reviewer/kb1 --reviewer-llm1-model openai/gpt-5-mini --reviewer-llm2-model openai/gpt-5-mini --out-dir output/reviewer/out_models

Source: main.py:parse_args, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer, openrouter_llm.py:LLMConfig

Example: reviewer with per-role endpoint overrides:

python main.py --mode reviewer --kb2-dir input/reviewer/kb2 --reviewer-llm1-base-url http://localhost:1234/v1 --reviewer-llm2-base-url http://localhost:1234/v1 --out-dir output/reviewer/out_local_roles

Source: main.py:parse_args, autogenbook/pipelines/reviewer_pipeline.py:run_reviewer, openrouter_llm.py:OpenRouterLLM.__init__

Cross-mode advanced flags

Flag	Default	Notes
--fail-fast-schema	false	Fail immediately on schema validation errors (also sets AUTOGENBOOK_FAIL_FAST_SCHEMA=1). (main.py:main, autogenbook/agents/base.py:BaseAgent._validate_with_repair)
--llm-base-url	OpenRouter base URL	Override OpenAI-compatible LLM endpoint for the run (LM Studio, local gateway, etc.). (main.py:parse_args, openrouter_llm.py:OpenRouterLLM.__init__)

Environment variables (complete)

OpenRouter

Variable	Required	Behavior	Source
OPENROUTER_API_KEY	Yes (OpenRouter)	Required for OpenRouter; optional for local endpoints.	openrouter_llm.py:OpenRouterLLM.__init__
OPENROUTER_HTTP_REFERER	No	Optional header.	openrouter_llm.py:OpenRouterLLM.__init__
OPENROUTER_X_TITLE	No	Optional header.	openrouter_llm.py:OpenRouterLLM.__init__
OPENROUTER_INPUT_COST_PER_M	No	Cost estimation input rate.	openrouter_llm.py:OpenRouterLLM.__init__, main.py
OPENROUTER_OUTPUT_COST_PER_M	No	Cost estimation output rate.	openrouter_llm.py:OpenRouterLLM.__init__, main.py
OPENROUTER_MAX_RETRIES	No	Retry count for transient errors (default 3).	openrouter_llm.py:_chat_once
OPENAI_API_KEY	No	Optional fallback API key for OpenAI-compatible servers.	openrouter_llm.py:OpenRouterLLM.__init__

AutoGenBook general

Variable	Required	Behavior	Source
AUTOGENBOOK_LLM_BASE_URL	No	Override OpenAI-compatible base URL for all LLM calls.	openrouter_llm.py:OpenRouterLLM.__init__
AUTOGENBOOK_LLM_API_KEY	No	Optional API key override for non-OpenRouter endpoints.	openrouter_llm.py:OpenRouterLLM.__init__
AUTOGENBOOK_FORCE_MINI_MODEL	No	Forces openai/gpt-5-mini for all runs.	openrouter_llm.py:OpenRouterLLM.__init__
AUTOGENBOOK_FAIL_FAST_SCHEMA	No	Fail immediately on schema validation.	main.py:main, autogenbook/agents/base.py:BaseAgent._validate_with_repair
AUTOGENBOOK_NONINTERACTIVE	No	Skip interactive prompts.	autogenbook/pipelines/book_pipeline.py:_ask_choice, autogenbook/pipelines/proposal_pipeline.py:_is_noninteractive
AUTOGENBOOK_ASSUME_YES	No	Auto-accept yes/no prompts (book).	autogenbook/pipelines/book_pipeline.py:_ask_yes_no
AUTOGENBOOK_SMOKE_FAST	No	Fast smoke mode.	autogenbook/smoke_test.py:main

Knowledge base

Variable	Required	Behavior	Source
AUTOGENBOOK_KB_OCR	No	Enable OCR for PDFs.	rag_kb.py:KnowledgeBase.build_from_directory
AUTOGENBOOK_KB_OCR_LANG	No	OCR language (default eng).	rag_kb.py:KnowledgeBase.build_from_directory
AUTOGENBOOK_KB_HEADING_CHUNKS	No	Chunk Markdown by headings.	rag_kb.py:KnowledgeBase.build_from_directory

MCP retrieval cache

Variable	Required	Behavior	Source
AUTOGENBOOK_MCP_CACHE_DIR	No	Cache directory for MCP tool results.	autogenbook/retrieval/mcp_papers.py:MCPPaperRetriever.__post_init__
AUTOGENBOOK_MCP_CACHE_TTL_S	No	Cache TTL (seconds).	autogenbook/retrieval/mcp_papers.py:MCPPaperRetriever.__post_init__
AUTOGENBOOK_MCP_CACHE_MAX_FILES	No	Cache size limit.	autogenbook/retrieval/mcp_papers.py:MCPPaperRetriever.__post_init__

Proposal

Variable	Required	Behavior	Source
AUTOGENBOOK_PROPOSAL_LLM1_MODEL .. AUTOGENBOOK_PROPOSAL_LLM5_MODEL	No	Per-role model overrides.	autogenbook/pipelines/proposal_pipeline.py:_resolve_model_override
AUTOGENBOOK_PROPOSAL_LLM1_BASE_URL .. AUTOGENBOOK_PROPOSAL_LLM5_BASE_URL	No	Per-role base URL overrides.	autogenbook/pipelines/proposal_pipeline.py:run_proposal
AUTOGENBOOK_PROPOSAL_PREV_SECTIONS	No	Number of previous sections kept for context.	autogenbook/pipelines/proposal_pipeline.py:run_proposal
AUTOGENBOOK_LLM1_MAX_ROUNDS	No	Max LLM1 rounds.	autogenbook/pipelines/proposal_pipeline.py:run_proposal
AUTOGENBOOK_PROPOSAL_SECTION_CHUNK_CHARS	No	Section prompt chunk size.	autogenbook/pipelines/proposal_pipeline.py:run_proposal
AUTOGENBOOK_PROPOSAL_SECTION_MAX_CHUNKS	No	Max section chunks.	autogenbook/pipelines/proposal_pipeline.py:run_proposal
AUTOGENBOOK_PROPOSAL_REVIEW_CHUNK_CHARS	No	Review prompt chunk size.	autogenbook/pipelines/proposal_pipeline.py:run_proposal
AUTOGENBOOK_PROPOSAL_FINAL_REVIEW_REPAIRS	No	Final review repair attempts.	autogenbook/pipelines/proposal_pipeline.py:run_proposal

Reviewer

Variable	Required	Behavior	Source
AUTOGENBOOK_REVIEWER_LLM1_MODEL	No	Reviewer LLM1 override.	autogenbook/pipelines/reviewer_pipeline.py:_resolve_model_override
AUTOGENBOOK_REVIEWER_LLM2_MODEL	No	Reviewer LLM2 override.	autogenbook/pipelines/reviewer_pipeline.py:_resolve_model_override
AUTOGENBOOK_REVIEWER_LLM1_BASE_URL	No	Reviewer LLM1 base URL override.	autogenbook/pipelines/reviewer_pipeline.py:run_reviewer
AUTOGENBOOK_REVIEWER_LLM2_BASE_URL	No	Reviewer LLM2 base URL override.	autogenbook/pipelines/reviewer_pipeline.py:run_reviewer

Tavily

Variable	Required	Behavior	Source
TAVILY_API_KEY	No	Enables Tavily web search.	autogenbook/retrieval/tavily.py:search_web

MCP gateway

Variable	Required	Behavior	Source
MCP_GATEWAY_ENABLE	No	Enable/disable MCP gateway.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_URL	No	Base URL for gateway.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_TRANSPORT	No	sse or streaming.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_RPC_PATH	No	RPC path (default /mcp).	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_SSE_PATH	No	SSE path (default /sse).	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_MESSAGE_PATH	No	Message path (default /message).	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_RPC_URL	No	Full RPC URL override.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_SSE_URL	No	Full SSE URL override.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_MESSAGE_URL	No	Full message URL override.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_TIMEOUT	No	Gateway timeout seconds.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_CACHE_TTL	No	Tool list cache TTL seconds.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_SSE_TIMEOUT	No	SSE stream timeout.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_SSE_ENDPOINT_TIMEOUT	No	SSE endpoint discovery timeout.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_API_KEY	No	Gateway API key.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_API_HEADER	No	Header for API key.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_API_PREFIX	No	Header prefix.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_GATEWAY_PROMPT_ECHO	No	Echo API key prompt.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_PROTOCOL_VERSION_SSE	No	MCP protocol version for SSE.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_PROTOCOL_VERSION_HTTP	No	MCP protocol version for streamable HTTP.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_CLIENT_NAME	No	MCP client name.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_CLIENT_VERSION	No	MCP client version.	mcp_gateway.py:MCPGatewayClient.__init__
MCP_CLIENT_CAPABILITIES_JSON	No	JSON capabilities advertised to MCP.	mcp_gateway.py:MCPGatewayClient._client_capabilities

Python runtime

Variable	Required	Behavior	Source
PYTHONUTF8	No	Set to 1 at startup.	main.py
PYTHONIOENCODING	No	Set in smoke tests.	autogenbook/smoke_test.py:_run

Complete reference: docs/CONFIGURATION.md. (docs/CONFIGURATION.md)

Outputs and artifacts

Common artifacts for all modes:

• run_meta.json: Run metadata, configuration, and OpenRouter usage totals. (autogenbook/llm_usage.py:write_run_meta)
• llm_usage.jsonl: Per-call usage/cost log with modality breakdown. (autogenbook/llm_usage.py:log_usage)
• logs/run.log: Console + file logging when enabled. (autogenbook/logging.py:get_logger, autogenbook/paths.py:default_run_paths)

Book/paper structure:

• structure_graph.json: Serialized document graph. (autogenbook/graph/doc_graph.py:save_graph_json, autogenbook/state.py:RunContext)
• sections/*.md: Leaf node content (default). Legacy LaTeX uses sections/*.tex. (book_builder.py:generate_contents, autogenbook/pipelines/paper_pipeline.py:run_paper)

Book-specific:

• context_memory.json: Running terminology and continuity hints. (book_builder.py:generate_contents, autogenbook/memory/context_memory.py:ContextMemory)
• section_reviews/*.json: Reviewer feedback if enabled. (book_builder.py:generate_contents)

Paper-specific:

• related_work.json: Literature agent output. (autogenbook/pipelines/paper_pipeline.py:run_paper)
• refs.bib: BibTeX bibliography (BibTeX mode). (autogenbook/pipelines/paper_pipeline.py:run_paper)

Scientist-specific:

• experiments/<run_id>/metrics.json, stdout.log, stderr.log. (autogenbook/pipelines/scientist_pipeline.py:_run_experiment)
• review.json: Review agent output. (autogenbook/pipelines/scientist_pipeline.py:run_scientist)

Proposal-specific:

• outline/, draft/, final/: Intermediate and final proposal artifacts. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• sections/<section_dir>/section.md, citations.json, used_sources.json, tool_calls.json. (autogenbook/pipelines/proposal_pipeline.py:_write_section_artifacts)
• final/proposal_final.md: Final proposal. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)

Reviewer-specific:

• review_final.md: Review output. (autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)

Audit artifacts:

• audit_report.json when audits are enabled. (autogenbook/audit/latex_auditor.py:audit_latex)

Repository layout and customization

• prompts/<mode>/: Prompt packs loaded per mode via the loaders in autogenbook/prompts/*_loader.py. (autogenbook/prompts/book_loader.py:load_book_prompts, autogenbook/prompts/paper_loader.py:load_paper_prompts, autogenbook/prompts/presentation_loader.py:load_presentation_prompts, autogenbook/prompts/scientist_loader.py:load_scientist_prompts, autogenbook/prompts/proposal_loader.py:load_proposal_prompts, autogenbook/prompts/reviewer_loader.py:load_reviewer_prompts)
• autogenbook/templates/toy_classification/: Scientist experiment template cloned into output/experiments/<run_id>/. (autogenbook/pipelines/scientist_pipeline.py:_prepare_run_dir, autogenbook/templates/toy_classification/run_experiment.py:main)
• preamble/robustness.tex: LaTeX robustness preamble copied into outputs and included in generated documents. (utils.py:ensure_robustness_preamble, book_builder.py:build_latex_document, autogenbook/pipelines/paper_pipeline.py:_build_paper_latex)
• input/ and examples/: Sample TXT specs and KB inputs used by the example commands in this README. (input/book/book_input.txt, input/paper/paper_input.txt, input/proposal/proposal_input.txt, examples/book_input.txt)
• scripts/check_latex_log.py: Utility to scan LaTeX logs for common errors. (scripts/check_latex_log.py:main)

Prompt packs (how to modify + what each prompt does)

Prompt packs are plain Markdown files under prompts/<mode>/ and are loaded at runtime by the mode-specific loaders (autogenbook/prompts/*_loader.py). Each loader enumerates required filenames in DEFAULT_*_PROMPT_FILES; missing files raise errors on startup. (autogenbook/prompts/book_loader.py:DEFAULT_BOOK_PROMPT_FILES, autogenbook/prompts/paper_loader.py:DEFAULT_PAPER_PROMPT_FILES, autogenbook/prompts/presentation_loader.py:DEFAULT_PRESENTATION_PROMPT_FILES, autogenbook/prompts/scientist_loader.py:DEFAULT_SCIENTIST_PROMPT_FILES, autogenbook/prompts/proposal_loader.py:DEFAULT_PROPOSAL_PROMPT_FILES, autogenbook/prompts/reviewer_loader.py:DEFAULT_REVIEWER_PROMPT_FILES)

How to modify safely:

• Edit the relevant file in prompts/<mode>/ and keep placeholder tokens. autogenbook/prompts/agent_prompts.py:render injects {GLOBAL_SYSTEM_POLICY} and {GLOBAL_EVIDENCE_INSTRUCTIONS} when referenced. (autogenbook/prompts/agent_prompts.py:render)
• Proposal prompts use {{LANGUAGE}} placeholders that are replaced in autogenbook/pipelines/proposal_pipeline.py:_apply_language. (autogenbook/pipelines/proposal_pipeline.py:_apply_language)
• For proposal prompt formatting rules and delimiters, see prompts/proposal/README.md. (prompts/proposal/README.md)
• After edits, run python -m autogenbook.smoke_prompts to check prompt completeness. (autogenbook/smoke_prompts.py)

Book mode prompts (prompts/book/)

Markdown-first uses the _md.md prompt variants when present (e.g., book_section_writer_*_md.md). LaTeX prompts remain available for --legacy-tex. (autogenbook/prompts/book_loader.py:load_book_prompts)

Prompt file	Main idea / purpose
prompts/book/global_system_policy.md	Global grounding rules and style constraints for book agents.
prompts/book/global_evidence_instructions.md	Evidence/citation rules for retrieved excerpts (RID/cite_key).
prompts/book/json_only_system.md	Force JSON-only output when a JSON response is required.
prompts/book/json_repair_system.md	Repair invalid JSON output and return only corrected JSON.
prompts/book/mcp_tools_system.md	Tool-use guidance for MCP paper/search tools.
prompts/book/book_json_from_txt_system.md	System role for converting a TXT spec into book JSON.
prompts/book/book_json_from_txt_user.md	User template defining the book JSON schema and input fields.
prompts/book/book_redundancy_system.md	System role for redundancy review of a book outline JSON.
prompts/book/book_redundancy_user.md	User task: detect redundant topics and propose a cleaned outline.
prompts/book/structure_subdivider_system.md	System role for subdividing outline nodes.
prompts/book/structure_subdivider_user.md	User task: split a parent section into children with page budgets.
prompts/book/book_section_writer_system.md	System role for book section writing.
prompts/book/book_section_writer_user.md	User task: write a single book section LaTeX body with grounding.
prompts/book/book_section_reviewer_system.md	System role for reviewing book sections (grounding/consistency).
prompts/book/book_section_reviewer_user.md	User task: produce review JSON with issues and required fixes.
prompts/book/book_section_revision_system.md	System role for revising sections based on review feedback.
prompts/book/book_section_revision_user.md	User task: apply required fixes to LaTeX body.
prompts/book/context_memory_system.md	System role for maintaining cross-section terminology memory.
prompts/book/context_memory_user.md	User task: update context memory JSON from new section text.
prompts/book/length_control_system.md	System role for length-adjusting LaTeX without changing meaning.
prompts/book/length_control_user.md	User task: expand/condense a section to target length.

Paper mode prompts (prompts/paper/)

Markdown-first uses the _md.md prompt variants when present (e.g., paper_section_writer_*_md.md). LaTeX prompts remain available for --legacy-tex. (autogenbook/prompts/paper_loader.py:load_paper_prompts)

Prompt file	Main idea / purpose
prompts/paper/global_system_policy.md	Global grounding rules and style constraints for paper agents.
prompts/paper/global_evidence_instructions.md	Evidence/citation rules for retrieved excerpts (RID/cite_key).
prompts/paper/json_only_system.md	Force JSON-only output when a JSON response is required.
prompts/paper/json_repair_system.md	Repair invalid JSON output and return only corrected JSON.
prompts/paper/mcp_tools_system.md	Tool-use guidance for MCP paper/search tools.
prompts/paper/paper_json_from_txt_system.md	System role for converting a TXT spec into paper JSON.
prompts/paper/paper_json_from_txt_user.md	User template defining the paper JSON schema and input fields.
prompts/paper/structure_subdivider_system.md	System role for subdividing outline nodes.
prompts/paper/structure_subdivider_user.md	User task: split a parent section into children with page budgets.
prompts/paper/paper_section_writer_system.md	System role for paper section writing.
prompts/paper/paper_section_writer_user.md	User task: write a single paper section LaTeX body with grounding.
prompts/paper/web_literature_system.md	Literature agent: summarize related work and novelty risks as JSON.
prompts/paper/web_literature_user.md	User wrapper for the literature agent’s input prompt.
prompts/paper/length_control_system.md	System role for length-adjusting LaTeX without changing meaning.
prompts/paper/length_control_user.md	User task: expand/condense a section to target length.

Presentation mode prompts (prompts/presentation/)

Prompt file	Purpose
prompts/presentation/global_system_policy.md	Global grounding rules and style constraints for presentation agents.
prompts/presentation/global_evidence_instructions.md	Evidence/source rules for retrieved excerpts (RID/cite_key).
prompts/presentation/json_only_system.md	Force JSON-only output when a JSON response is required.
prompts/presentation/json_repair_system.md	Repair invalid JSON output and return only corrected JSON.
prompts/presentation/mcp_tools_system.md	Tool-use guidance for MCP paper/search tools.
prompts/presentation/presentation_json_from_txt_system.md	System role for converting a TXT spec into presentation JSON.
prompts/presentation/presentation_json_from_txt_user.md	User template defining the presentation JSON schema and inputs.
prompts/presentation/structure_subdivider_system.md	System role for subdividing slide groups.
prompts/presentation/structure_subdivider_user.md	User task: split a parent slide group into child slides.
prompts/presentation/presentation_slide_writer_system.md	System role for slide body writing in Markdown.
prompts/presentation/presentation_slide_writer_user.md	User task: write a single slide Markdown body with grounding.
prompts/presentation/presentation_narration_system.md	System role for slide narration generation.
prompts/presentation/presentation_narration_user.md	User prompt template for slide narration generation.
prompts/presentation/presentation_image_prompt_system.md	System guidance for slide image prompt generation.
prompts/presentation/presentation_image_prompt_user.md	Image prompt template using slide text and prior summary.
prompts/presentation/length_control_system.md	System role for length-adjusting slide Markdown without changing meaning.
prompts/presentation/length_control_user.md	User task: expand/condense slide body to target length.

Scientist mode prompts (prompts/scientist/)

Prompt file	Main idea / purpose
prompts/scientist/global_system_policy.md	Global grounding rules and style constraints for scientist agents.
prompts/scientist/global_evidence_instructions.md	Evidence/citation rules for retrieved excerpts (RID/cite_key).
prompts/scientist/json_only_system.md	Force JSON-only output when a JSON response is required.
prompts/scientist/json_repair_system.md	Repair invalid JSON output and return only corrected JSON.
prompts/scientist/mcp_tools_system.md	Tool-use guidance for MCP paper/search tools.
prompts/scientist/idea_agent_system.md	System role for generating candidate research ideas.
prompts/scientist/idea_agent_user.md	User task: propose multiple testable ideas with hypotheses.
prompts/scientist/literature_agent_system.md	System role for novelty/related-work analysis.
prompts/scientist/literature_agent_user.md	User task: assess novelty and build related-work scaffold.
prompts/scientist/plan_agent_system.md	System role for selecting an idea and building a plan.
prompts/scientist/plan_agent_user.md	User task: output experiment plan + paper outline JSON.
prompts/scientist/code_patch_agent_system.md	System role for safe code patch proposals.
prompts/scientist/code_patch_agent_user.md	User task: propose unified diff to improve the metric.
prompts/scientist/analyze_agent_system.md	System role for interpreting run artifacts into claims.
prompts/scientist/analyze_agent_user.md	User task: summarize results and propose figures/tables.
prompts/scientist/review_agent_system.md	System role for structured peer review of the draft.
prompts/scientist/review_agent_user.md	User task: produce review JSON with issues and scores.
prompts/scientist/revision_agent_system.md	System role for applying review-driven revisions.
prompts/scientist/revision_agent_user.md	User task: revise LaTeX text to address review issues.
prompts/scientist/paper_section_writer_system.md	System role for LaTeX section writing in scientist mode.
prompts/scientist/paper_section_writer_user.md	User task: write one paper section LaTeX body.
prompts/scientist/length_control_system.md	System role for length-adjusting LaTeX without changing meaning.
prompts/scientist/length_control_user.md	User task: expand/condense a section to target length.

Proposal mode prompts (prompts/proposal/)

Prompt file	Main idea / purpose
prompts/proposal/llm1_architect_system.md	LLM1 (architect): define proposal meta-prompt and compliance blueprint.
prompts/proposal/llm2_researcher_outline_system.md	LLM2 (researcher outline): design proposal structure and plan.
prompts/proposal/llm3_opponent_outline_system.md	LLM3 (opponent outline): strict review of outline/compliance.
prompts/proposal/llm4_researcher_writer_system.md	LLM4 (researcher writer): draft full proposal sections.
prompts/proposal/llm5_opponent_final_system.md	LLM5 (opponent final): final compliance/quality review.
prompts/proposal/metadata_extractor_system.md	Extract proposal metadata from the input text.
prompts/proposal/json_only_system.md	Force JSON-only output when a JSON response is required.
prompts/proposal/json_repair_system.md	Repair invalid JSON output and return only corrected JSON.
prompts/proposal/mcp_tools_system.md	Tool-use guidance for MCP paper/search tools.

Reviewer mode prompts (prompts/reviewer/)

Prompt file	Main idea / purpose
prompts/reviewer/prompt1_default.md	Default LLM1 “architect” prompt when KB1 is unavailable. (autogenbook/pipelines/reviewer_pipeline.py:run_reviewer)
prompts/reviewer/architect_prompt.md	LLM1 instruction to synthesize a final reviewer system prompt from KB1.
prompts/reviewer/prompt2.md	LLM2 review prompt template (rubric/structure for the review).
prompts/reviewer/reviewer_system.md	Generic reviewer system guidance (structured, evidence-grounded).
prompts/reviewer/reviewer_user.md	Generic reviewer user template for review inputs.

Retrieval, citations, and audits

• Local KB indexing uses BM25; chunks include stable rid and cite_key identifiers. (rag_kb.py:KnowledgeBase, rag_kb.py:Chunk)
• Retrieval items are normalized into RetrievalItem objects. (autogenbook/retrieval/types.py:RetrievalItem)
• LaTeX citations are extracted and normalized; BibTeX entries are produced from retrieval items when needed. (autogenbook/citations/extract.py:extract_citations, autogenbook/citations/ledger.py:CitationLedger)
• LaTeX auditing flags unknown citations, missing figures, and numeric claims without evidence. (autogenbook/audit/latex_auditor.py:audit_latex)

Operations & troubleshooting (short)

• If lualatex is missing, use --no-pdf. (book_builder.py:compile_pdf, main.py:parse_args)
• Proposal mode requires MCP paper tools; it errors if tools are unavailable. (autogenbook/pipelines/proposal_pipeline.py:run_proposal)
• Reviewer mode requires a non-empty KB2. (autogenbook/pipelines/reviewer_pipeline.py:run_reviewer, tests/test_reviewer_mode.py)
• Scan LaTeX logs with python scripts/check_latex_log.py <file.log> to catch common errors early. (scripts/check_latex_log.py:main)

Full details: docs/OPERATIONS.md and docs/TROUBLESHOOTING.md. (docs/OPERATIONS.md, docs/TROUBLESHOOTING.md)

Developer quickstart

• Install deps: pip install -r requirements.txt. (requirements.txt)
• Run tests: python -m unittest. (tests/test_reviewer_mode.py)
• Prompt pack smoke tests: python -m autogenbook.smoke_prompts. (autogenbook/smoke_prompts.py)
• Schema smoke tests: python -m autogenbook.schemas.smoke. (autogenbook/schemas/smoke.py)
• End-to-end smoke: python -m autogenbook.smoke_test (requires API key unless AUTOGENBOOK_SMOKE_FAST=1). (autogenbook/smoke_test.py)

More details: docs/DEVELOPER_GUIDE.md. (docs/DEVELOPER_GUIDE.md)

Security notes

• Secrets are read from environment variables only. (openrouter_llm.py:OpenRouterLLM.__init__, mcp_gateway.py:MCPGatewayClient.__init__)
• Retrieved context is sanitized to reduce prompt-injection patterns. (autogenbook/retrieval/sanitize.py:sanitize_context_text)
• Scientist patching enforces safety rules for diffs and file paths. (autogenbook/runner/patch_apply.py:_safety_scan_diff, autogenbook/runner/patch_apply.py:_sanitize_patch_path)

More details: docs/SECURITY.md. (docs/SECURITY.md)

Docs map

See docs/INDEX.md for the full documentation index. (docs/INDEX.md)

Limitations

• CLI only; no HTTP server is defined in the codebase. (main.py:parse_args, autogenbook/orchestrator.py:run)
• LLM calls require credentials for the configured endpoint (OpenRouter requires OPENROUTER_API_KEY). (openrouter_llm.py:OpenRouterLLM.__init__)
• PDF generation requires LuaLaTeX. (book_builder.py:compile_pdf)

License / Contributing

License and contribution guidelines are not defined in runtime code paths; confirm in the repo root before distribution. (Needs confirmation.)