This section assumes familiarity with mathematical programming.Documentation Index
Fetch the complete documentation index at: https://docs.feasible.cloud/llms.txt
Use this file to discover all available pages before exploring further.
Capabilities matrix
| Class | Status |
|---|---|
| LP (continuous linear) | Solid |
| MILP (mixed-integer linear) | Solid — primary use case |
| QP / MIQP (quadratic) | Solid for convex objectives |
| NLP (continuous nonlinear) | Works for smooth, well-conditioned problems |
| MINLP (mixed-integer nonlinear) | Works; convergence depends on the relaxation |
| Conic (SOCP / SDP-shaped) | Works |
What works well
- Problems an experienced modeler could write in 30–100 lines of JuMP. Variable counts in the hundreds, constraint counts in the low thousands, well-described mathematically.
- Problems with clear domain semantics. “One queen per region” is easy. “Inventory at end of week equals inventory at start plus production minus demand” is easy. The clearer the prose, the cleaner the formulation.
- Re-formulations. If you want to try a different relaxation or alternative formulation, asking the agent to redo it is fast.
- Pre-computed coefficient tables. Time-indexed costs, demand patterns, distance matrices. The agent will inline these as data and keep the model linear when possible.
What is rough
- Very large or sparse problems. The agent has to reason about every variable in context — this falls apart past a few hundred variables. For industrial-size problems, hand-write the model and use Feasible only for the human-friendly answer formatting.
- Subtle nonlinearities. Bilinear terms, logical implications encoded as products of binaries. The agent often gets these almost right and needs a nudge.
- Numerical conditioning. Big-M values, scaling. The agent picks reasonable defaults but won’t tune them automatically. If your problem is numerically delicate, expect to specify scales explicitly.
- Disjunctive constraints. Indicator / disjunction modeling is supported but error-prone. If your problem has many disjunctions, double-check the formulation in the Constraints panel.
- Stochastic / robust formulations. Out of scope. Feasible models the deterministic version of whatever you describe.
Worked example: chemical reactor (NLP)
A small nonlinear problem that exercises the agent on
exp(·) terms and shows what the Constraints panel looks like for nonlinear expressions.