Retrieval Radar

An AI discovery engine that reads real Google Photos complaints and sorts them by how retrieval broke — not just how people felt.

84 tagged 705 analysed 3 public sources

Ask the corpus

84 tagged complaints indexed

Try

Source

View

Where retrieval breaks

Complaints per failure mode. Severity is rated 1–5; opportunity is volume × severity.

Corpus

Severity spread

Where the pain surfaces

The taxonomy

Every complaint is tagged against these five failure modes.

The memory–index mismatch

The clues people bring versus the clues the system demands. Where these two lists don’t meet is where retrieval fails. Select any clue to read the complaints behind it.

What people remember

demands↔recall

What search demands

How the engine works — and what it can’t tell you

1 Collect — 705 public complaints, 3 sources 2 Filter — keyword pre-pass for retrieval language 3 Classify — an LLM tags each complaint to the taxonomy + severity 4 Compare — rank by volume × severity 5 Interrogate — ask a question, the findings recompute on the matching subset

Limits, stated plainly. 84 of 705 complaints were judged retrieval-relevant, so every figure here rests on n=84 — enough to rank failure modes, not to size a market. 89% of the retrieval signal came from Reddit, where people describe what they were trying to find; store reviews rarely go past one line. Some complaints describe photos that went missing after a sync or restore rather than photos that exist but can’t be surfaced — the two look alike in a keyword pass, and separating them is the next pass worth doing. Ask the corpus matches complaints by term overlap against the question (expanded with a retrieval-domain synonym map); it retrieves and re-aggregates real complaints rather than generating prose, so nothing on screen is invented.