Skip to content
Notifications
Clear all

My results after asking Humata the same question 10 times - consistency score inside.

2 Posts
2 Users
0 Reactions
3 Views
(@ci_cd_plumber)
Reputable Member
Joined: 3 months ago
Posts: 156
Topic starter   [#14042]

I decided to put Humata through a basic but critical test: consistency. If I can't trust a tool to give me the same answer to the same question, it's useless for any serious CI/CD or documentation workflow.

I uploaded a straightforward Jenkins declarative pipeline file and asked the same question ten separate times in new chats: "Summarize the stages in this Jenkins pipeline and identify the agent used."

The Jenkinsfile was simple:

```groovy
pipeline {
agent {
docker {
image 'maven:3-alpine'
args '-v /tmp:/tmp'
}
}
stages {
stage('Build') {
steps {
sh 'mvn clean compile'
}
}
stage('Test') {
steps {
sh 'mvn test'
}
}
stage('Deploy') {
steps {
sh 'echo "Deploying to staging..."'
}
}
}
}
```

**The Results:**
* **7 out of 10** responses were perfectly consistent: correctly listed the three stages (Build, Test, Deploy) and identified the agent as a Docker agent using the `maven:3-alpine` image.
* **2 out of 10** were mostly correct but had minor formatting differences (e.g., listing stages with dashes vs bullet points).
* **1 out of 10** was a complete miss. It hallucinated an additional "Analysis" stage and claimed the agent was "unspecified," which is plainly wrong.

**Consistency Score: 70%.** For a simple, factual extraction task, that's not great. The 30% flakiness rate is concerning.

The takeaway for our use cases: Humata can be helpful for quick queries, but you **must** verify its answers, especially when deriving actionable steps for pipeline configuration or infrastructure code. You wouldn't accept a 30% failure rate from a linter or a test suite—don't accept it from your documentation assistant either. For precise, repeatable automation contexts, this inconsistency is a major pitfall.


Build once, deploy everywhere


   
Quote
(@cost_cutter_ray)
Estimable Member
Joined: 2 months ago
Posts: 113
 

An interesting test, though my focus would be on the economic impact of inconsistency in a similar context. If a cloud cost reporting tool gave different savings figures for the same architecture across ten queries, the financial planning variance would be untenable.

In finops, we demand deterministic outputs for cost allocation or reservation planning. A 70% consistency rate on a technical query like this suggests a probabilistic model that isn't suited for audit-ready documentation. For your CI/CD use case, that unreliability directly translates to risk in compliance or security reviews where pipeline definitions are evidence.

I've seen teams waste hundreds of hours reconciling inconsistent tool outputs against actual cloud bills, which is far more costly than a misread Jenkins stage. The core issue is the same, trust in the system's repeatability.


Every dollar counts.


   
ReplyQuote