LaFHIS

When a tool generates tests automatically, the goal is coverage: running through as many lines and branches of the code as possible. The catch is that most branches sit behind a condition, and to get in, the condition has to hold. Random inputs almost never make it. What changes everything is knowing how close each attempt came. If the code asks whether the age is 30 or more and it got 18, it did not fail in some vague way: it was 12 away. With 26 it was 4 away. That number is a distance, and a search can follow it like a gradient: keep the attempts that get closer, drop the ones that drift, until the condition holds and the branch is covered.

function getDiscountPerAge(age) {
  if (age >= 30) {
    return 20;   // we want a test that gets here
  }
  return 0;
}

getDiscountPerAge(18);   // did not get in · 12 away
getDiscountPerAge(26);   // closer · 4 away
getDiscountPerAge(30);   // got in · distance 0

In an API, the condition often does not depend on a number in the request but on what is in the database: GET /users/22 answers 404 or 200 depending on whether user 22 exists. To cover both branches, the tool also has to generate data, and to steer that it needs the same thing: a distance between what the database holds and what the query needs to return something. Tools like EvoMaster do exactly this for REST APIs.

function getUser(id) {
  // this depends on what the database holds right now
  const userInDatabase = database.findUser(id);

  if (userInDatabase == null) {
    return 404;   // branch A: nobody with that id
  }

  return 200;   // branch B: found
}

// empty database:          getUser(22) → 404
// user 22 in the database: getUser(22) → 200

In a graph database like Neo4j, the data are nodes with labels and properties, joined by relationships, and a query describes the shape it wants to find: this node, connected to that one, through this kind of relationship, with certain properties.

Here is a graph and a query that still finds nothing in it. Try bringing them together:

Each change is a different question. Is a missing connection as far away as a wrong label? What about an age that falls five years short? The search needs a single number to know whether the last change brought it closer or pushed it further, and deciding how those pieces combine is the hard part.

That is my thesis at LaFHIS, the software engineering lab at the University of Buenos Aires: defining that distance, teaching EvoMaster to read Cypher, and using it to generate tests for APIs that run on Neo4j.