Understanding quantum oracles
An oracle, $O$, is an unexposed operation that is used as input to another algorithm. Often, such operations are defined using a classical function $f : \{0, 1\}^n \to \{0, 1\}^m$, which takes an $n$-bit binary input and produces an $m$-bit binary output. To do so, consider a particular binary input $x = (x_{0}, x_{1}, \dots, x_{n-1})$. You can label qubit states as $\ket{\vec{x}} = \ket{x_{0}} \otimes \ket{x_{1}} \otimes \cdots \otimes \ket{x_{n-1}}$.
You might first attempt to define $O$ so that $O\ket{x} = \ket{f(x)}$, but this method has a couple of problems. First, $f$ might have a different size of input and output ($n \ne m$), such that applying $O$ would change the number of qubits in the register. Second, even if $n = m$, the function might not be invertible: if $f(x) = f(y)$ for some $x \ne y$, then $O\ket{x} = O\ket{y}$ but $O^\dagger O\ket{x} \ne O^\dagger O\ket{y}$. This means you can't construct the adjoint operation $O^\dagger$, and oracles have to have an adjoint defined for them.
Define an oracle by its effect on computational basis states
You can deal with both of these problems by introducing a second register of $m$ qubits to hold the answer. Then, define the effect of the oracle on all computational basis states: for all $x \in \{0, 1\}^n$ and $y \in \{0, 1\}^m$,
$$ \begin{align} O(\ket{x} \otimes \ket{y}) = \ket{x} \otimes \ket{y \oplus f(x)}. \end{align} $$
Now $O = O^\dagger$ by construction and you resolved both of the earlier problems.
Tip
To see that $O = O^{\dagger}$, note that $O^2 = \mathbf{1}$ since $a \oplus b \oplus b = a$ for all $a, b \in {0, 1}$. As a result, $O \ket{x} \ket{y \oplus f(x)} = \ket{x} \ket{y \oplus f(x) \oplus f(x)} = \ket{x} \ket{y}$.
Importantly, defining an oracle this way for each computational basis state $\ket{x}\ket{y}$ also defines how $O$ acts for any other state. This behavior follows immediately from the fact that $O$, like all quantum operations, is linear in the state that it acts on. Consider the Hadamard operation, for example, which is defined $H \ket{0} = \ket{+}$ and $H \ket{1} = \ket{-}$. If you wish to know how $H$ acts on $\ket{+}$, you can use that $H$ is linear,
$$ \begin{align} H\ket{+} & = \frac{1}{\sqrt{2}} H(\ket{0} + \ket{1}) = \frac{1}{\sqrt{2}} (H\ket{0} + H\ket{1}) \\ & = \frac{1}{\sqrt{2}} (\ket{+} + \ket{-}) = \frac12 (\ket{0} + \ket{1} + \ket{0} - \ket{1}) = \ket{0}. \end{align} $$
When defining the oracle $O$, you can similarly use that any state $\ket{\psi}$ on $n + m$ qubits can be written as
$$ \begin{align} \ket{\psi} & = \sum_{x \in \{0, 1\}^n, y \in \{0, 1\}^m} \alpha(x, y) \ket{x} \ket{y}. \end{align} $$
Here, $\alpha : \{0, 1\}^n \times \{0, 1\}^m \to \mathbb{C}$ represents the coefficients of the state $\ket{\psi}$. Thus,
$$ \begin{align} O \ket{\psi} & = O \sum_{x \in \{0, 1\}^n, y \in \{0, 1\}^m} \alpha(x, y) \ket{x} \ket{y} \\ & = \sum_{x \in \{0, 1\}^n, y \in \{0, 1\}^m} \alpha(x, y) O \ket{x} \ket{y} \\ & = \sum_{x \in \{0, 1\}^n, y \in \{0, 1\}^m} \alpha(x, y) \ket{x} \ket{y \oplus f(x)}. \end{align} $$
Phase oracles
Alternatively, you can encode $f$ into an oracle $O$ by applying a phase based on the input to $O$. For example, you might define $O$ such that $$ \begin{align} O \ket{x} = (-1)^{f(x)} \ket{x}. \end{align} $$
If a phase oracle acts on a register initially in a computational basis state $\ket{x}$, then this phase is a global phase and hence not observable. But such an oracle can be a powerful resource if applied to a superposition or as a controlled operation. For example, consider a phase oracle $O_f$ for a single-qubit function $f$. Then, $$ \begin{align} O_f \ket{+} & = O_f (\ket{0} + \ket{1}) / \sqrt{2} \\ & = ((-1)^{f(0)} \ket{0} + (-1)^{f(1)} \ket{1}) / \sqrt{2} \\ & = (-1)^{f(0)} (\ket{0} + (-1)^{f(1) - f(0)} \ket{1}) / \sqrt{2} \\ & = (-1)^{f(0)} Z^{f(0) - f(1)} \ket{+}. \end{align} $$
Note
Note that $Z^{-1}=Z^{\dagger}=Z$ and therefore $Z^{f(0)-f(1)}=Z^{f(1)-f(0)}.$
More generally, both views of oracles can be broadened to represent classical functions, which return real numbers instead of only a single bit.
Choosing the best way to implement an oracle depends heavily on how this oracle is to be used within a given algorithm. For example, Deutsch-Jozsa algorithm relies on the oracle implemented in the first way, while Grover's algorithm relies on the oracle implemented in the second way.
For more information, see the discussion in Gilyén 1711.00465.