Definability is most often considered only for a first-order language $L$ (not a theory); a definable set is an equivalence class of formulas? which evaluate in the same way in every $L$-structure. It makes sense to do the same for a theory: in that case in the definition of the equivalence relation one restricts to models of $T$.
Let $L$ be a first-order language and $T$ a theory in $L$. Any formula $\phi(x_1,\ldots,x_n)$, whose only free variables could be among $x_1,\ldots,x_n$, together with a model $M$ for $T$ evaluates to a truth value on each $a = (a_1,\ldots,a_n)\in M^n$, hence it determines the subset $\phi(M)\subset M^n$ of all $a$ such that $\phi(a)$. Two formulas $\phi,\psi$ with the same number of free variables are equivalent if $\phi(M) = \psi(M)$ for every model $M$ of $T$. A definable set $X$ in $T$ is an equivalence class of formulas in $L$ under this relation. Consider the category $\mathcal{M}_{el}(L)$ of $L$-structures and elementary monomorphism?s and its full subcategory $\mathcal{M}_{el}(T)$ whose objects are the models of $T$. We can also view a definable set for a theory $T$ as a functor $\mathcal{M}_{el}(T)\to Set$ which is of the form $M\mapsto\phi(M)$ for some $L$-formula $\phi$.
By $X(M)$ denote the set of all $a\in M$ such that $X(a)$ holds, i.e. $\phi(a)$ is true for any choice of representative $\phi\in X$.
Given two definable sets $X,Y$ in $T$ a definable function $f: X\to Y$ is a definable subset of the product sort $X\times Y$ such that $f_M\in X(M)\times Y(M)$ is a function (we also write $f_M : X(M)\to Y(M)$) for every model $M$ of $T$. Analogously a definable equivalence relation is a definable subset $R\subset X\times X$ such that $R(M)$ is an equivalence relation for every model $M$ of $T$.
Proposition. Given a small category $\mathcal{M}$ and two functors $F,G:\mathcal{M}\to Set$ the natural transformations $\alpha : F\to G$ are in 1-1 correspondence with functors $\tilde\alpha : \mathcal{M}\to Set$ such that $\tilde\alpha(A) \subset F(A)\times G(A)$ and such that $\tilde\alpha(A)$ is a functional relation.
Proof. If $\alpha:F\to G$ is an $Ob(\mathcal{M})$-indexed family with components $\alpha_A: F(A)\to G(A)$ viewed as functional relations $\tilde\alpha_A\subset F(A)\times G(A)$, then the extension to a functor means
that there is a (automatically unique) dotted arrow $\tilde\alpha_f$
making the diagram commutative (once the existence of $\tilde\alpha_f$ is known, the functoriality of $\tilde\alpha_{f\circ g} = \tilde\alpha_f\circ\tilde\alpha_g$ comes by uniqueness of $\tilde\alpha_f$ and the functoriality of $F\times G$ which it restricts from). That means that for every $x\in F(A)$ $(x,\alpha_A(x))$ should be sent to $(F(f)(x), G(f)(\alpha_A(x)))$ by $\tilde\alpha_f$, what is a restriction of $F(f)\times G(f)$, but by $\alpha_B$ being functional relation this must be the same as $(F(f)(x), \alpha_B(F(f)(x)))$, i.e. that $G(f)(\alpha_A(x))=\alpha_B(F(f)(x))$. Q.E.D.
In other words, functoriality of $\tilde\alpha$ is the same as $\alpha$ being natural.
Corollary. Definable functions for $L$ (for $T$) are in 1-1 correspondence with natural transformations of definable sets as functors $\mathcal{M}_{el}(L)\to Set$ (resp. $\mathcal{M}_{el}(T)\to Set$).
For a fixed (multi-sorted, first-order) theory $T$, definable sets and definable functions form a category $Def(T)$. This category (or any equivalent category) is the syntactic category of the theory. Models of $T$ can be recovered as left exact functors $Def(T)\to Set$. Notice that definable sets are functors from models to sets, and models are functors definable sets to sets; just the latter are the “evaluation functionals” among all functors.
See also definable groupoid.
An $\infty$-definable set is an intersection of definable sets.
We can also consider a relative version of a definable set (definability with parameters).
Given definable sets $X,Y$, a fixed model $M$ and a fixed set $A\subset X(M)$, we say that an element $b\in Y(M)$ is definable over $A$ if there is $(a_1,\ldots,a_n)\in A^n$ and a $T$-definable function $f:X^n\to Y$ such that $f(a_1,\ldots,a_n)=b$. All elements $b$ definable over $A$ form the subset $Y(A)\subset Y(M)$. A subset $B\subset M$ is definably closed if it is closed under definable functions. Given $A\subset M$, there is the minimal definably closed subset $B$ such that $A\subset B\subset M$, the definable closure $B = dcl(A)$ of $A$.
If we extend the language by the constants from $A$ we get a new language $L_A$. The definable sets in language $L$ over $A$ are the definable sets in language $L_A$ over $\emptyset$.
One can also study ind-objects and pro-objects in the category of definable sets and definable functions.
Moshe Kamensky, Ind- and Pro- definable sets, math.LO/0608163
Jan Holly, Definable operations on sets and elimination of imaginaries, Proc. Amer. Math. Soc. 117 (1993), no. 4, 1149–1157, MR93e:03052, doi, pdf
David Kazhdan, Lecture notes in model theory, pdf
D. Haskell, E. Hrushovski, H.D.Macpherson, Definable sets in algebraically closed valued fields: elimination of imaginaries, J. reine und angewandte Mathematik 597 (2006), MR2007i:03040, doi/CRELLE.2006.066)