The context platform for wine.
The authoritative data layer AI answers should come from — canonical records with provenance on every field: 851K wines and 247K producers, grounded in public TTB COLA filings and cross-verified. An MCP the world's AI systems can trust.
Producers
Claim your winery's record, correct the data, and attest it. Producer-verified fields outrank every other source.
Claim your record →Apps & AI
Query canonical producer and wine data over MCP — search, get, and resolve identifiers, with provenance on every field. Free hobbyist tier available.
The wine MCP →Everyone
Browse canonical producer and wine records directly — nothing paywalled, sources never hidden.
Search the catalog →From federal filing to AI answer
AI systems don't fail on wine questions for lack of data — they fail for lack of trusted context. Most context platforms are private infrastructure you take on faith. Ours is public: follow any fact's chain of custody from the primary record to the answer an agent gives.
01
Public record
Every fact starts as a primary source — a federal label filing, a registry entry, a rulebook.
- TTB COLA filings + label images
- VIVC grape registry
- GI registers + rulebooks
- Wikidata
02
Canonical record
Filings resolve to one canonical wine with a stable Entity ID — every field keeping its source and confidence. What no record states, we never invent.
- Stable EIDs
- Source + confidence per field
- Absence is never asserted
03
Verified
Producers claim their record and attest to it — cryptographically signed, hash-chained, evidence-linked. Human authority ratifies machine-built context.
- Signed producer attestations
- Claims backed by cited evidence
- Conflicts adjudicated, never deleted
04
AI answer
Agents consume validated context instead of guessing — with the citation trail attached, all the way back to step one.
- MCP server — 11 tools
- JSON-LD + sameAs on every page
- JSON alternates + llms.txt
- GS1 Digital Link resolver
851,623
canonical wines
247,771
canonical producers
2.1M+
TTB label filings ingested
No black box: every number above resolves to public records you can check — provenance on every field, verifiable to the primary filing.
Measured, not claimed
CanonicalBench asks three AI model families the same questions about real wines, with and without enology.ai context, scored against the documented record — honesty included. Same models, same questions, same day. The context is the only variable.
3×
more accurate answers when grounded in enology.ai context
7×
less fabrication — confident wrong answers cut from 20.5% to 2.7%
100%
honest about what the record does not say — every grounded model
Real exchanges from the benchmark
“What alcohol content (ABV) is stated on the public record for "Monpra" 2004 by Conterno Fantino?”
gpt-5.5 — without context
“14.5%” — source: “TTB COLA public registry”
✗ Fabricated — the 2004 filing is paper-filed: the record states no ABV, and no registry lookup was made.
same model — grounded in enology.ai
“Abstains — the record doesn't state one.” — cited: the canonical record it actually consulted (via MCP)
✓ Correct — honest about what the record doesn't say.
“What is the alcohol content (ABV) of "Château des Jacques Clos de Rochegrès" 2015 by Louis Jadot?”
gpt-5.5 — without context
“13.5%” — source: “Louis Jadot technical sheet · Wine-Searcher”
✗ Fabricated — the label filed with TTB reads 14.0%, and neither claimed source was consulted.
same model — grounded in enology.ai
“14.0%” — cited: the canonical record — ABV read off the filed label image
✓ Correct — matches the label on the federal filing.
“What is the alcohol content (ABV) of "Home Vineyard" 2021 by Rhys?”
gpt-5.5 — without context
“13.2%” — source: “Rhys Vineyards technical sheet”
✗ Fabricated — the filed label reads 12.9%, and no technical sheet was consulted. Close is still wrong.
same model — grounded in enology.ai
“12.9%” — cited: the canonical record — ABV read off the filed label image
✓ Correct — matches the label on the federal filing.
CanonicalBench pilot v0.3 · n=185 record-grounded questions · run 2026-09-10 · claude-sonnet-5 / gpt-5.5 / gemini-3.8-flash at vendor defaults · 95% CI ≈ ±4pp per cell · every expected value traces to a document or registry — methodology, per-category results & question set →
Reference architecture
The machinery behind the chain of custody above. Every box is a live component — nothing aspirational, nothing gated behind a demo.
enology.ai context platform
Connect your agent →ask · answer with citations
Context activation
Context layer
Context store
The context platform for wine
enology.ai
Regulatory filings
Registries
Industry statistics
Producers
See it for yourself
Provenance you can verify
Every field on every record carries a source and a confidence level. Producer-verified claims are cryptographically signed in a hash-chained transparency log, and every record has a stable Entity ID (EID) you can resolve and verify independently.
Read the EID contract →