REGISTRY VS. RAG · BENCHMARKS
Ingest once. Query forever.
Our benchmark against retrieval on an enterprise corpus: accuracy on questions that need the whole corpus, and cost per correct answer.
- CORPUS
- 28 filings; ranking, set and count questions
- ARMS
- Talonic fast path, Talonic loop, RAG hybrid, RAG long-context
- HEADLINE
- asked which of 28 filings is biggest, RAG got 0 of 5 right
- REPRODUCIBLE
- frozen corpus manifest, query set and price table
- FULL METHOD
- talonic.com/vs/rag
Same corpus, same questions.
| Measure | Talonic fast path | RAG hybrid | RAG long-context |
|---|---|---|---|
| Ranking questions correct (n = 24) | 83.3% | 12.5% | 37.5% |
| Set questions correct (n = 25) | 84.0% | 12.0% | 48.0% |
| Entity recall, mean | 93.1% | 34.5% | 63.7% |
| Model cost per query | $0.0015 | $0.16 | $0.32 |
| Model cost per correct ranking answer | $0.0018 | $1.26 | $0.85 |
| One-time cost per document, paid once | $0.17 | $0 | $0 |
Talonic fast path, RAG hybrid and RAG long-context, one run each, same corpus and query set. The Talonic loop arm, the ten-run spreads for both RAG arms, latency and the full method are on the benchmark page. RAG arms are fully reproducible scripts with a pinned price table; the Talonic arm runs the platform API end to end.
RAG re-reads. The registry already read.
Retrieval-augmented generation re-reads your documents on every question, and it can only re-read the passages it retrieved. A question that needs the whole corpus, which of these contracts is biggest, how many carry an escalation clause, which ones expire this quarter, is a question retrieval cannot see.
Our registry structures every document once, at higher accuracy, and answers every question after that from the database. The same question returns the same rows, every row keeps the line and the scan region that produced it, and you pay to read a document once, not once per question per user.
Cost that stops growing.
- Deterministic and cited. The same question returns the same rows, and every row keeps the line and the scan region that produced it. A RAG answer changes with the retrieval; ours changes when the document does.
- A fraction of the tokens. You pay to read a document once, at $0.17 in this run. After that a question costs $0.0015 in model time, against $0.16 to $0.32 for a RAG arm that re-reads every time, and a great deal more per correct answer.
- The curve crosses early. Retrieval is cheaper on the first question and more expensive on every one after. For a corpus that a finance team asks about daily, the crossover is weeks, not years.
What this benchmark does not show.
- It is deliberately narrow. Ranking, set and count questions across a whole corpus. Retrieval is fine for a single-passage lookup, and this run does not claim otherwise.
- Per-field precision is not on this page. Extraction precision is measured per customer, per schema, and published on their case page when they release it. The benchmark measures question answering, not field extraction.
- One corpus, one query set. Twenty-eight filings and one fixed query set. Enough to show the mechanism; not a survey. The bundle is available on request so you can run it on your own corpus.
The registry the benchmark runs on is the architecture described in DIN SPEC 91491, Europe’s first standard for AI-ready data, co-authored by Talonic with Fraunhofer IIS, Humboldt-Innovation, GIIC and the German standards body.
Run it on your own corpus.
The corpus manifest, query set, retrieval configurations and price table are frozen and available as a bundle. Bring your own documents and we run the same four arms against them.
Note: Every figure on this page is lifted from the registry benchmark at /vs/rag, same run, same arms, same price table. Nothing here is re-measured or rounded differently.