AI Glossary / Graphs and symbolic AI / Intermediate

Procedural attachment

Definition

Procedural attachment connects a symbol or query in a knowledge-based reasoning system to an executable procedure. When the relevant expression is encountered under suitable conditions, the system can call the procedure to calculate a value, retrieve information, or test a relationship. It provides a bridge between symbolic reasoning and specialized computation or external data.

What it means

A reasoning system may know rules about quantities without deriving every arithmetic result from elementary logical axioms. It may also need information held in a database rather than copied into its knowledge base. An attached procedure supplies such results through a defined interface.
For example, a function representing a geometric distance can be connected to a numerical calculation. A relation representing an inventory record can be connected to a lookup. The procedure must implement the intended meaning of the symbol. Attaching arbitrary code does not automatically make its output a justified logical fact, and different systems impose different requirements on when attachments may run.

How it works

Define the logical symbol and the procedure's contract: required inputs, returned outputs, units, possible failures, and the interpretation of each result. The reasoner invokes the attachment when the needed arguments are available. It translates the result into the form used by its reasoning process and continues with the surrounding rules.
A procedure may require fully specified inputs even though the logical relation can be written with variables. That operational restriction matters during search. A failed network request, an unknown answer, and a confirmed false statement should not be silently treated as equivalent. For repeatable reasoning, record the relevant data version or time whenever an external answer can change.

Three practical examples

A symbolic configuration checker can call a numerical routine to test a calculated capacity against a modeled requirement.
A question-answering system can consult an external catalog for an item's attributes while combining them with rules about categories.
A robot task planner can invoke a geometric feasibility procedure when a high-level action requires a physically possible motion.

Think of it this way

Think of a planner who asks a specialist calculator for a measurement and then uses that measurement in a written decision rule. The calculator fills one gap; it does not replace the whole argument. The analogy is limited because software interfaces must define exactly which input combinations are supported and how uncertain, unavailable, or erroneous results are represented.

Worked example

Consider an invented layout checker with two fixed points: A = (0, 0) and B = (3, 4), measured in meters. A symbolic rule permits a connection when the Euclidean distance between endpoints is at most the specified cable length. The distance function is attached to a numerical procedure implementing sqrt((x2 − x1)^2 + (y2 − y1)^2).
With these inputs, the procedure returns sqrt(9 + 16) = 5 meters. A cable length of 5.5 meters therefore satisfies the modeled rule, while 4.9 meters does not. The surrounding reasoner can combine this result with separate facts about connector compatibility.
Now suppose the caller supplies B = (300, 400) in centimeters without converting units. Treating those numbers as meters produces 500 meters and an incorrect rejection. The arithmetic is internally consistent, but the attachment's input contract has been violated. Also, the example only checks straight-line length; obstacles, slack, and routing are absent from the invented model and cannot be inferred from this calculation.

Illustrative case study

Fictional case: Copper Shelf builds a catalog assistant that reasons about which storage boxes fit a compartment. Its rules use an attached dimensions lookup and a small comparison routine. During a test, a missing product record is incorrectly converted to “does not fit.”
The team changes the interface to report “dimensions unavailable” separately from a failed fit test. It also records the catalog version used in each answer. This fictional story illustrates why the meaning of a procedure's result is as important as obtaining it. It claims no measured savings or real deployment success.

Common misconceptions and mistakes

Do not assume an attachment supports every logical direction. A routine that computes distance from coordinates does not automatically enumerate coordinates at a requested distance. Do not treat successful code execution as proof that its assumptions match the knowledge model.
Hidden side effects can make repeated reasoning calls alter the world or change later answers. Stale caches can preserve facts after their sources change. Numerical approximations require care near decision boundaries. If a procedure reports a heuristic result rather than a verified answer, the reasoner should preserve that distinction instead of presenting it as an unconditional theorem.

Business use

Procedural attachment can make a knowledge-based application useful without translating every established calculator or database into logical axioms. The business question is whether the combined answer remains understandable and reproducible. A reviewer should be able to see which conclusions came from rules and which depended on an external calculation or record.
For a pilot, choose a small attachment with a clear contract and test ordinary inputs, missing data, unit mismatches, and boundary values. Monitor latency if reasoning may call the procedure many times. Treat changes to the attached routine as changes to the application's effective knowledge and behavior.

Check your understanding

In the layout example, does a 5-meter distance result prove that a 5.5-meter cable can be installed successfully?

Answer: It proves only that the straight-line length condition is satisfied under the stated coordinates and units. Successful installation also depends on any additional conditions, such as connectors and routing, that the model must represent separately.

Related terms

Knowledge representation
Theorem proving
Ground term
external procedure
symbolic reasoning
interface contract
database query
task and motion planning
soundness

Learn next

Learn logical relations and grounding next. Write an attachment contract for the distance routine, including units and unsupported inputs. Then decide how the system should represent an unavailable coordinate, a calculation error, and a valid distance that exceeds the allowed length.

Sources

Cleo Condoravdi and colleagues, Natural Language Access to Data: It Takes Common Sense!, original SRI paper describing SNARK procedural attachment: https://www.sri.com/wp-content/uploads/2021/12/2006.pdf

Previous: Problem-solving agent Next: Procedural memory