Benchmarks

Our retrieval is the differentiator.

On multi-hop questions — where the answer is split across several documents — Tableside finds the answer 92%of the time, where classic keyword search manages 77%. Here’s the public benchmark that proves it, run side by side with the search most chatbots ship by default.

The hard test — answers split across documents

MultiHop-RAG

Public benchmark — Tang & Yang 2024

2,556 questions whose answer is split across two to four different articles. Ask “is the tandoori roti gluten-free and under £4?” and ordinary search finds one fact and misses the other. This is the real test of whether a system can read across documents.

92%

of these questions, answered

Tableside finds the answer 92% of the time — 15 points ahead of the keyword search most apps still rely on.

Scored on all 2,255 of the benchmark’s questions, using its own official answer key.

Tableside

reads across documents to connect the facts

92%

Standard AI search

embedding similarity — what most RAG apps use

78%

Keyword search

classic keyword matching

77%

Share of questions where the answer was found — same benchmark, same scoring for every system.

96%

Connecting facts

piece together clues across articles

93%

Comparing things

weigh two reports against each other

84%

Tracking over time

what changed between two dates

Curious how it works, how it compares to other approaches, and where it still slips up? It’s all in the technical whitepaper.

Every number here is measured on the public MultiHop-RAG benchmark using its own official scoring — not a private test we graded ourselves. The full method, the head-to-head against other retrieval approaches, and an honest look at where it still slips up are in the technical whitepaper.