Agentic RAG Design Configurator

A multi-agent system that turns natural-language design requests (plus optional reference images) into a parametric propeller mesh + 3 renders + a quality-check report, with optional post-session persistence of each agent's reflections into an embedding-ready database for later RAG.

Built on LangChain with swappable LLM backends (OpenAI / Anthropic / Google). The prompt and tool architecture is DC-agnostic by design — propeller-specific knowledge lives in DC_prompt_fragments/ and can be swapped for another design configurator.

How it works

Nine stateful sub-agents collaborate via a flat horizontal dispatcher (eight in-session, one post-session).

Default flow (PLANNER_FIRST=False, UII-first):

user → Receptionist → Orchestrator → User Input Inspector
                   → Planner → DC Input Creator → [DC Input Inspector]
                   → Tool Caller (mesh + renders)
                   → DC Output Inspector → Orchestrator → Receptionist → user

Any agent can ESCALATE to the Orchestrator, which calls the Planner for a Problem/Solution/Sequence plan and re-routes one step at a time. On REVISE, a new attempt folder is opened.

Post-session save (opt-in): when the user types quit, the system asks whether to save the session to the database. If yes, the Database Handler interviews each in-session agent through a per-field ASK:/SAVE: protocol and writes one .txt file per scheduled field, shaped to be embedding-ready (self-contained, declarative, one topic per file, ≤700 cl100k_base tokens for Semantic fields).

Project layout

.
├── main.py                       # entrypoint → agents.loader.run()
├── config.py                     # paths + RhinoCompute env vars
├── requirements.txt
├── workflow_settings/settings.py # 11 runtime flags (see Configuration)
├── agents/
│   ├── loader.py                 # session lifecycle, REPL, archival, DH invocation
│   ├── step_caps.py              # single source of truth for every MAX_*
│   ├── shared/                   # prompt assembly, routing, retry, rate-limit, tools
│   ├── orchestrator/  receptionist/  planner/  user_input_inspector/
│   ├── dc_input_creator/  dc_input_inspector/  tool_caller/  dc_output_inspector/
│   └── database_handler/         # post-session interviewer (opt-in)
├── DC_prompt_fragments/          # DC-specific prompt fragments (propeller today)
│   ├── dc_config/                # parameters, structure, capabilities, constraints
│   └── tools_config/             # tool inventory, render-check library
├── tools/
│   ├── generate_mesh/            # RhinoCompute + Grasshopper definition
│   ├── render_mesh/              # trimesh + pyvista backends
│   └── calculate/
└── extra_utilities/              # TODO_known_issues.md, warnings_developer.md, smoke tests

Generated at runtime (gitignored): attempts/, logs/, previous_sessions/, database/, inputs/.

Setup

Requires Python 3.13 and a running RhinoCompute instance for mesh generation.

git clone https://github.com/R-SMP/agentic-rag-design-config.git
cd agentic-rag-design-config

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Environment variables

Copy each .env.example to .env and fill in the values:

copy .env.example .env
copy agents\.env.example agents\.env

File	Purpose
.env (root)	RHINO_COMPUTE_URL, RHINO_COMPUTE_API_KEY
agents/.env	Default LLM_PROVIDER, MODEL_NAME, and the matching API key (OPENAI_API_KEY / ANTHROPIC_API_KEY / GOOGLE_API_KEY)
agents/<agent>/.env	Optional per-agent override (falls back to agents/.env, then root .env)

⚠️ Real .env files are gitignored — never commit API keys.

Running

python main.py

A REPL opens. Type a design request (and optionally drop reference images into inputs/input_images/ paired with <name>_note.txt files). Type quit to end the session; the system will ask whether to save to the database.

Configuration

All runtime flags live in workflow_settings/settings.py:

Flag	Default	Notes
MESH_CHECKS	False	watertight / volume / degenerate-face checks
RENDER_LIBRARY	"trimesh"	"trimesh" or "pyvista" metric backend
RAG_ENABLED	False	reserved (logged but unwired)
DC_INSPECTOR_ENABLED	True	run DCII before mesh generation
CHAIN_ACCESS	True	Orchestrator sees inter-agent chain messages
KEEP_IMAGES_IN_CONTEXT	False	image bytes persist across hand-offs
RATE_LIMIT_ENABLED	True	throttle every llm.invoke()
RATE_LIMIT_REQUESTS_PER_SECOND	1.0	steady-state call rate
DCOI_COMPARISON_MODE	3	1=user inputs, 2=extraction, 3=both
PLANNER_FIRST	False	True = Planner runs before UII
EMBEDDING_MODEL	"text-embedding-3-large"	for DH-shaped Semantic bodies
EMBEDDING_VECTOR_DIMS	1024	MRL truncation dim at index time
EMBEDDING_MAX_RESPONSE_TOKENS	700	DH cap for Semantic bodies

Status & known issues

• extra_utilities/TODO_known_issues.md — open issues (O1–O8) and carry-forward bugs.
• extra_utilities/warnings_developer.md — load-bearing invariants (W1–W12) that must not regress.

Roadmap

This repo (test11) is the v5 starting point. Near-term direction:

• Move heavy compute (RhinoCompute, mesh gen, rendering) off the local driver onto a server.
• Build a web interface as the user-facing front-end.
• Wire RAG_ENABLED to consume the database the Database Handler is now producing.