Tools · Runs in your browser, optional AI deep-dive
AI Architecture Picker
A RAG assistant, an AI agent, a document pipeline, a voice agent, a vision system or a copilot in your product? Pick what you're building, answer a few questions, and get a complete system architecture: an interactive diagram, the key decisions and why, a suggested stack, a build roadmap and the risks to plan for.
Want this built, or reviewed?
I design and build production AI systems end-to-end — RAG assistants, agents and automations, document pipelines, and AI features inside existing products. If you have a design already, I can review it against your real data and constraints.
Frequently asked questions
Which kinds of AI system does this cover?
Six: knowledge assistants that answer from your documents (RAG), AI agents and workflow automation, document and data extraction, voice agents, vision (images, video and live cameras), and copilots or content generation inside your product. Each has its own questions, and each result is a full system diagram with the decisions, a suggested stack, a three-phase roadmap and the risks to plan for.
Should I build an AI agent or a fixed workflow?
If you can write the steps down, build a workflow: code decides the order and the model handles the fuzzy parts — classifying, extracting, drafting. It is cheaper, more predictable and easier to debug. Reach for an agent that plans its own steps only when the steps genuinely depend on what it finds along the way.
When should I not use RAG at all?
When the whole corpus fits comfortably in the model's context window and doesn't change by the minute — a single handbook or a few policy documents. Putting the documents in the prompt, with prompt caching, is simpler and avoids retrieval misses. RAG earns its complexity once the corpus is too big or too costly to send on every request.
What is the difference between RAG, hybrid RAG and GraphRAG?
Standard RAG retrieves passages by embedding similarity. Hybrid RAG adds a keyword (BM25) index so exact terms like product codes and error messages are found too, and merges both result lists with a reranker. GraphRAG also extracts entities and relationships into a knowledge graph, so it can answer questions that connect several facts across documents.
Is pgvector good enough for production?
Yes, for most teams. Below a few million chunks pgvector performs well and keeps vectors in the same Postgres database as the rest of your data, with transactions, backups and SQL filtering. Dedicated engines like Qdrant or Pinecone pay off at tens of millions of vectors or very high query volume.
Can I build RAG without sending data to OpenAI or other cloud APIs?
Yes. Open-weight LLMs and embedding models can be served inside your own network with tools like Ollama or vLLM, with pgvector or self-hosted Qdrant for retrieval. The trade-off is running GPUs yourself and evaluating which open model answers your questions well enough.
Does this tool store my answers?
No. The recommendation is computed in your browser.
What does the AI deep-dive add?
It turns the recommendation into a design for your actual system — for example chunking and evaluation questions for a knowledge assistant, tool permissions and approval points for an agent, a field schema and validation rules for extraction, or a latency budget for a voice agent. The optional AI deep-dive sends your answers and description to an AI model through OpenRouter and returns a written review. It needs a free sign-in (email code) and is rate-limited. Nothing you paste is saved by this site — only a count of runs, for the limits.