RAG cannot rank a corpus. We measured it.

RAG answers from the documents it retrieves. For a lookup question that is fine: the answer lives in one document, and retrieval only has to find it. A ranking question is different. To know which of 28 companies has the highest revenue, a system has to have read all 28. A top-k retrieval window sees a handful and guesses from those.

Talonic captures every filing once at ingest into a canonical data plane: typed fields with provenance, resolved against a field registry. The ranking question then becomes an ordered scan over captured fields across every company. Ingest is paid once. Every query after that runs over structured data.

RAG paid $0.11 in model cost for these five questions, measured from API usage tokens at pinned rates, and pays it again on every repeat. Talonic pays at ingest, once per document. Every ranking query after that runs over already-structured data.

Read the table honestly: the long-context steelman is competitive on single-document numeric reads. It sees everything, at $0.32 per query and six-figure token counts per call, and it still fails the majority of corpus-ranking and set questions. Hybrid retrieval fails nearly all of them. The structural gap is in questions that require having read the whole corpus.

The cost rows are measured, both ways, and they price two different things. $1.05 is the natural-language agent: a multi-step tool loop (287 model calls across these 39 queries) that plans, reads structured fields, and writes an answer. The data-plane query underneath it is an API read over fields captured once at ingest, with roughly zero marginal model cost: programmatic consumers pay $0.17 per document once and then query for free. RAG has no such split. Its per-query price is its only query path, re-paid on every question, forever, over unstructured text.

Divide by correctness and the picture inverts. On ranking questions, Talonic costs $1.05 per query at 87.5% accuracy: $1.20 per correct answer. The long-context arm costs $0.32 at 33.3%: $2.88 per correct answer. The hybrid arm is cheap per attempt and wrong five times out of six. And this 53-filing corpus is the most favorable size the long-context arm will ever see: its cost grows linearly with corpus size until the corpus no longer fits at all, while a query over structured fields does not re-read the corpus.

• Corpus: SEC 10-K financial statements. Headline needle: 28 latest filings, one per company, fiscal year 2024. Full battery: 53 filings across the same 28 companies.
• Gold: SEC EDGAR XBRL frames, the numbers each company itself filed. Deterministic scoring, values scale-normalized at 1% tolerance to the gold magnitude band. No LLM judges.
• Parsing: identical OCR-to-markdown (Mistral OCR) fed to both arms in the needle benchmark, so parse quality is not the variable.
• Models: Claude Sonnet 4.6 for both needle arms and both battery RAG arms. The battery Talonic column is the production agent over the public API (run 7, 2026-08-18), which routes across models internally.
• RAG configs: needle arm BM25, top-k 12 chunks, one model call. Battery hybrid arm: BM25 + dense embeddings, reciprocal rank fusion, top-24 chunks. Long-context arm: whole-corpus partitioned full-context calls with merge.
• Costs: measured, never estimated. RAG arms: API usage tokens at pinned rates. Talonic: the platform's metered model usage for this corpus and query run (2026-08-18), split into ingestion (one-time, $9.06 for 53 documents, all pipeline stages included) and agent queries ($41.05 for 39 queries).
• Preregistration: protocol, corpus manifest, query set, configs, and price table frozen before any query ran. The RAG arms are fully reproducible from the benchmark scripts; the Talonic arm is auditable end to end.