[Bug]: AI Models catalog falls back to three models when estimated_pulls is null

[mikeshollen]

Hi all.

First, thank everyone that is involved with this project. It's epic! This is my first time reporting a bug or even using Github, but here it goes.

[Bug]: AI Models catalog falls back to three models when estimated_pulls is null

Description

The AI Models page in Project NOMAD displays only three models (llama3.1, deepseek-r1, and llama3.2), even though the public model catalog is accessible and contains hundreds of models.

Clicking "Refresh Models" does not resolve the issue.

I traced the problem to a null-handling bug in the backend's model-sorting logic, identified the offending code, and successfully restored the full GUI catalog with a minimal two-line patch.

Environment

• OS: Lubuntu 26.04.1
• Application: Project NOMAD, current version as of October 25, 2026
• Docker image: ghcr.io/crosstalk-solutions/project-nomad:latest
• Container: nomad_admin
• Ollama: 0.40.11
• Browser: Firefox

Steps to Reproduce

1. Navigate to Settings → AI Assistant → AI Models.
2. Attempt to refresh the downloadable model catalog.
3. Observe that only llama3.1, deepseek-r1, and llama3.2 are displayed.
4. Query the local models API with popularity sorting:

curl -sS 'http://localhost:8080/api/ollama/models?sort=pulls&recommendedOnly=false&limit=15&force=true'

The endpoint returns only three fallback models.

5. Change the sorting parameter to alphabetical:

curl -sS 'http://localhost:8080/api/ollama/models?sort=name&recommendedOnly=false&limit=15'

This successfully returns 15 models, with hasMore: true.

Root Cause

The public model catalog includes entries where:

"estimated_pulls": null

In admin/app/services/ollama_service.ts, the sortModels() function calls parsePulls() without checking whether the input is a string.

The compiled JavaScript contains:

const parsePulls = (pulls) => {
    const multiplier = pulls.endsWith('K')

When pulls is null, the function throws:

Cannot read properties of null (reading 'endsWith')

The Command Center logs confirm the sequence:

[OllamaService] Fetching fresh available models from API
[OllamaService] Successfully cached available models
[OllamaService] Failed to retrieve models from Nomad API: Cannot read properties of null (reading 'endsWith')
[OllamaService] Returning fallback recommended models due to failure in fetching available models

The catalog downloads successfully, but processing fails during sorting.

Verified Local Fix

I modified the compiled JavaScript at:

/app/app/services/ollama_service.js

Added the following guard to parsePulls():

if (typeof pulls !== 'string') return 0;

I also added the equivalent defensive check to parseSize(), which has the same potential null-handling vulnerability:

if (typeof size !== 'string') return 0;

After validating JavaScript syntax, deploying the patched file, and restarting nomad_admin, the full model catalog was restored in the GUI.

Expected Behavior

Models with missing popularity statistics should remain available in the catalog.

Missing estimated_pulls values should be handled gracefully rather than causing the entire catalog request to fail.

Actual Behavior

The backend throws an exception and substitutes three fallback models, making the full downloadable catalog inaccessible through the GUI.

Suggested Permanent Fix

Add null/type guards to the sorting functions in the TypeScript source:

const parsePulls = (pulls: string | null | undefined): number => {
    if (typeof pulls !== 'string') return 0;
    // Existing logic
}

Apply equivalent handling to parseSize() for missing or null size values.

Ideally, add regression tests covering catalog entries with null popularity and size fields.

Impact

This bug can prevent users from browsing and downloading models through the standard GUI despite having a functioning internet connection and a valid model catalog.

The failure also misleadingly resembles a connectivity or cache problem, potentially encouraging unnecessary updates or reinstalls.

Status: Root cause identified; local workaround tested successfully. Permanent upstream fix requested.