From 400 Errors to Reliable Results: Lessons Configuring AI Search

The first result said 400. The second result was the same. Then everything clicked.

If you are wiring FileFortress AI search to a local OpenAI-compatible runtime (like Foundry), the fastest path to success is understanding the difference between:

Provider/setup failures (endpoint, model, compatibility)
Prompt intent failures (query does not map to search filters)

Most Common Root Cause

A model name that looks right but does not exist on the configured endpoint. The request can be valid JSON and still fail with 400.

Lesson 1: Model IDs Must Match Exactly

In many local runtimes, model IDs include backend/runtime suffixes. A generic model alias can fail.

# Looks plausible, may fail
model = phi-4

# Exact endpoint model ID, works
model = phi-4-openvino-gpu:1

That single difference can convert repeated 400 responses into successful translations.

Lesson 2: Empty-Body 400 Does Not Mean \"No Clue\"

Some providers return 400 with no response body. It still usually points to one of these:

Model not available on endpoint
Unsupported request parameter
Endpoint/path mismatch

A practical checklist:

Confirm endpoint health (foundry service status).
Confirm model IDs from /v1/models.
Set exact model with tools configure foundry --model ....
Re-test prompt with --dry-run --explain.

Lesson 3: Some Queries Are Not Search Queries

Once model setup was fixed, one prompt still failed:

\"How many remotes do i have?\"

That is a valid product question, but not a file-filter request. The AI command is focused on mapping prompts to search criteria, so it correctly returned \"no filters extracted\".

Command Selection Matters

Use ai for file-search intent.
Use remotes list for remote inventory and count.

Lesson 4: Dry-Run and Explain Are Your Fastest Debug Tools

These options turn debugging from guesswork into a deterministic loop:

# Validate interpretation only
filefortress ai "find images smaller than 5kb" --dry-run

# Inspect reasoning
filefortress ai "find images smaller than 5kb" --dry-run --explain

Once interpretation looks correct, execute without dry-run.

A Repeatable Recovery Workflow

Fix provider correctness (endpoint + exact model ID).
Validate with small, obvious prompt.
Use dry-run + explain for each new prompt shape.
Switch to explicit search flags when determinism is required.