DeckForge

Agentic Presentation Intelligence System Report-First | No Free Facts | Bilingual Output

DeckForge is an RFP-to-Deck engine built for Strategic Gears Consulting. It transforms institutional knowledge — stored across SharePoint presentations, proposals, reports, and frameworks — into consulting-grade proposal decks through a conversational, multi-agent workflow.

The system receives structured RFP summaries, searches and synthesizes relevant knowledge from a pre-indexed SharePoint corpus, generates a fully-cited research report for human approval, and converts that approved report into a branded slide deck in Arabic, English, or bilingual format.

---

Architecture Summary

Pipeline: 10 Steps, 5 Human Gates

1. RFP Intake — BD Station pushes structured AI Assist summary (10-field contract)
2. Context Understanding — Context Agent parses RFP into structured object → Gate 1
3. SharePoint Retrieval — Retrieval Agent executes 5-strategy search → Gate 2
4. Deep Analysis — Analysis Agent extracts atomic Claim Objects into Reference Index
5. Research Report — Research Agent writes fully-cited report → Gate 3 (most important)
6. Slide Outline — Structure Agent converts report to slide structure → Gate 4
7. Slide Content — Content Agent distills report into slide copy
8. Quality Assurance — QA Agent enforces No Free Facts, validates every claim
9. PPTX Rendering — Design Agent renders branded PPTX via python-pptx → Gate 5
10. Export — Final PPTX, research report (.docx), source index, gap report

9 Agents, 3 Models

Agent	Model	Role
Workflow Controller	LangGraph (deterministic)	State machine routing, gate enforcement
Conversation Manager	Claude Sonnet 4.6	Natural language → structured actions
Context Agent	GPT-5.4	RFP parsing into structured JSON
Retrieval Agent	GPT-5.4 + Azure AI Search	Query generation + source ranking
Analysis Agent	Claude Opus 4.6	Deep extraction into atomic Claim Objects
Research Agent	Claude Opus 4.6	Fully-cited research report generation
Structure Agent	GPT-5.4	Report → slide outline conversion
Content Agent	GPT-5.4	Slide copy writing from approved report
QA Agent	GPT-5.4	No Free Facts enforcement, validation
Design Agent	python-pptx (deterministic)	Branded PPTX rendering

Core Principles

• No Free Facts — every factual claim must trace to a cited source. Unsupported claims fail closed.
• Report-First — a research report is approved by humans before any slides are created.
• Gaps over guesses — missing evidence is flagged explicitly, never fabricated.

---

Current Milestone Status

Milestone	Scope	Status
M0: Scaffold + README	Project structure, config, dependencies, README	In progress
M1: Enums	src/models/enums.py — 19 StrEnum classes	Planned
M2: Pydantic Models	9 model files (common.py through indexing.py)	Planned
M3: Master State	state.py + ids.py + __init__.py re-export	Planned
M4: Config + LLM Wrapper	settings.py + models.py + llm.py	Planned
M5–M10: Agents, Pipeline, Rendering, Knowledge Layer	Full system	Deferred

---

Local Setup

Prerequisites

• Python 3.12+
• Git

Installation

git clone https://github.com/albarami/Deckbuilder.git
cd Deckbuilder

python -m venv .venv

# Windows
.venv\Scripts\activate

# macOS/Linux
source .venv/bin/activate

pip install -r requirements.txt

Environment

Copy .env.example to .env and fill in your API keys:

cp .env.example .env

Required keys: OPENAI_API_KEY, ANTHROPIC_API_KEY. Azure keys are needed for production only.

---

Test Commands

# Run all tests
.venv\Scripts\python.exe -m pytest tests/ -v

# Run tests for a specific module
.venv\Scripts\python.exe -m pytest tests/agents/test_enums.py -v

# Lint check
.venv\Scripts\python.exe -m ruff check src/

# Type check
.venv\Scripts\python.exe -m mypy src/

# All three (run before every review)
.venv\Scripts\python.exe -m pytest tests/ -v && .venv\Scripts\python.exe -m ruff check src/ && .venv\Scripts\python.exe -m mypy src/

---

Repo Workflow

Roles

• Cursor builds code
• Salim reviews, approves or rejects, and authorizes commits

Protocol

1. Read docs first — architecture, prompt library, state schema, existing code
2. Implement one thing — exactly what Salim asked, nothing more
3. Validate and test — pytest, ruff, mypy must all pass
4. Report results — state what was built, what was tested, what passed
5. Wait for approval — no commits until Salim says "commit"

Commit Convention

feat(scope): description    — new feature or agent
fix(scope): description     — bug fix
test(scope): description    — adding or updating tests
refactor(scope): description — code restructure (no behavior change)
docs(scope): description    — documentation changes
chore(scope): description   — config, dependencies, tooling

Rules

• Never commit without Salim's explicit approval
• Never push broken code
• Never force push
• Docs are the source of truth — if code contradicts docs, fix the code
• All Pydantic models in src/models/, never inline in agent files
• All LLM calls through src/services/llm.py, never direct API imports in agents

Remote

origin  https://github.com/albarami/Deckbuilder.git