Run the same question through five AI architectures — Plain, RAG, MCP, Agent, and Security-aware. Inspect every retrieval, tool call, and citation, then score groundedness, cost, and safety to pick the right one with evidence.
Same question, five architectures. See what each does under the hood, how the answer changes, and which one scores highest on groundedness, cost, and safety.
The model answers straight from its parametric memory. No retrieval, no tools.
Northwind Labs likely offers standard US PTO — around 15–20 days after several years of tenure. (Guessing — I have no access to their actual policy.)
Embed the question, pull the top-k matching chunks, stuff them into the prompt.
Northwind employees accrue 25 days of PTO per year after 2 years of employment. [policy-pto#0][policy-pto#1]
The model calls typed tools over a standard protocol — the same server works across clients.
Called `get_policy(topic="pto")` via MCP. Result: employees get 25 days per year after 2 years, 30 days after 5. [policy-pto]
Plan → act → observe → repeat. The model chooses tools and reacts to results across turns.
Searched policy docs, cross-referenced tenure brackets, then confirmed with `get_employee_handbook`. After 3 years an employee is in the 25-day bracket, plus 2 floating holidays. [policy-pto][handbook]
RAG over an OWASP/CWE corpus, plus tools that look up CWEs and scan code for common anti-patterns.
That query is vulnerable to CWE-89 (SQL injection) — user input is concatenated into SQL. Use parameterised queries (`db.query("… WHERE id=$1", [id])`). Maps to OWASP A03. [owasp-a03#0][cwe-89]
This is a frozen snapshot from a real run — same prompt, four architectures. Expand any View Trace to see exactly what happened.
No retrieval. Model answers from parametric memory only.
Embed → retrieve top-k chunks → stuff into prompt.
Model calls typed tools/resources over standard protocol.
Multi-step iterative loop: plan → act → observe → answer.
RAG over OWASP/CWE + security tools (lookup_cwe · scan_snippet).
Want to run your own? Sign up free — it takes 20 seconds.
Answer your team's questions against the bundled Northwind Labs handbook — five side-by-side pipelines, full traces, per-pipeline scores, and a recommended architecture you can export.
Free to try · No credit card · Your runs stay private to your account
Each approach optimises for different constraints. Use this as a quick cheat-sheet when deciding how to give a model the context it needs.
Creative writing, brainstorming, or anything general enough that the model already knows the answer.
Q&A over your own documents, policy lookups, product knowledge — anywhere grounding on a known corpus matters most.
Reusable tool/resource servers you want to share across apps, IDEs, or agents — the "USB-C" for context.
Multi-step tasks where the model must plan, call several tools, react to results, and iterate to a final answer.