TREC RAG 2026 Competition Architecture
One frozen authenticated source branches into three cache-first Retrieval submissions and two selected-evidence RAG submissions, with exact limits, run IDs, validation, and accepted status.
Final 2026 submission architecture and supporting reports.
Start with the accepted system: one frozen authenticated source, then Frozen source → Retrieval submissions + RAG submissions. Use the briefing, experiments, and 2025 writeups as supporting context rather than unfinished work.
1
Understand the final approachTrace the frozen source into three Retrieval runs and two RAG runs.
2
Verify the accepted filesUse the ledger for exact paths, run IDs, priorities, hashes, and validator links.
3
Browse supporting evidenceUse the briefing, experiment reports, and prior-year method library for context.
Available Reports
The final architecture is the canonical 2026 handoff. The briefing, merged experiment evidence, last year's method library, and historical architecture synthesis preserve supporting context.
Retrieval and RAGDoll analyses
These are one existing 119-topic Retrieval quality analysis and two separate 119-topic RAGDoll citation-support reports. RAGDoll is not an official TREC correctness score.
TREC RAG 2026 Briefing
Explains the competition, ClimbMix, sample documents, organizer nuggets, answer JSON, evaluation framing, and the original build strategy.
TREC RAG 2025 Writeups
A detailed interactive review of 15 team writeups, including reader paths, leaderboards, architecture panels, implementation recipes, and failure diagnosis.
Promising 2025 RAG Architecture
Narrative to verified answer: a source-backed synthesis of the strongest 2025 retrieval, coverage, evidence, and citation patterns.
Open architecture reportAll-topic tethered-facet validation
The portable canonical v3 report explains why RRF remains selected, where the DUAL arms regressed, and what the Topic 31/300 evidence supports next.
How To Read This Workspace
Pick a path based on what you need today. The pages are designed to cross-reference each other instead of repeating every detail.
Start with the final architecture
Read the whole-system map, candidate core, two RAG strategies, and accepted-artifact section for the completed project in one pass.
Audit the accepted submissions
Open the ledger for the exact five files, hashes, run identities, priority order, and repo-local validation workflow.
Read the preserved research
Use the briefing for task context and the 2025 reports as a method library. Historical experiments remain provenance, not a backlog.
Source Map
The reports keep generated explanation separate from source material, so readers can trace where each layer came from.
Local dev topics, nuggets, and qrels
Used by the briefing for sample query, sample ClimbMix doc IDs, organizer nuggets, and source caveats.
Proceedings PDFs and figures
Used by the writeup report for team methods, architecture panels, score caveats, and reusable playbooks.
Static checks for report integrity
Each interactive report has a Node smoke test that checks key sections, links, and safety constraints.