[Feature Request] Add a small "knowledge retrieval failure modes" note for STaRK users (docs only)

[onestardao]

Hi STaRK team,

Thank you for releasing STaRK. A benchmark that combines text and knowledge relations is very helpful for studying retrieval and reasoning together.

I have been working on compact failure-mode maps for retrieval and RAG systems and recently contributed a robustness-related entry to Harvard MIMS Lab’s ToolUniverse. When working with STaRK-like settings, I often see recurring issues:

• systems retrieve the right entities but miss key relations
• spurious relations are retrieved and used, leading to incorrect reasoning
• evaluation focuses on surface-form answers without checking relational grounding

I would like to propose a small, documentation-only note for STaRK users.

Proposed feature

Add a short markdown page under the docs, for example:

stark_knowledge_retrieval_failure_modes.md

The page could:

1. Describe typical failure modes in text-plus-relation retrieval.
2. For each mode, list:
   • example symptoms in STaRK evaluations
   • likely causes (indexing, representation, scoring)
   • simple diagnostics (e.g., inspecting retrieved relation sets).
3. Suggest a small checklist for error analysis:
   • whether the right entities were retrieved
   • whether the right relations were retrieved
   • where the reasoning step failed.

Motivation

• STaRK is a good testbed for understanding retrieval and relational reasoning together.
• A short failure-mode note would help users interpret errors with more structure and design better follow-up experiments.
• This is documentation only and can be kept concise.

If this seems useful and in scope, I would be happy to contribute a concise draft via PR.

Thank you for considering.