SME-level 5G NR answers where every claim is labeled grounded or inferred, and every citation is machine-verified against the spec archive before you ever see it. No vector database. No hallucinated references.
What triggers sr-TransMax exhaustion, and what does the UE do next?
When Scheduling Request transmissions reach sr-TransMax, the MAC entity ✓ grounded releases PUCCH/SRS, clears configured grants, and initiates Random Access on the SpCell.
Latency-sensitive traffic ~ inferred can then stall for the RA back-off.
✓ TS 38.321 §5.4.4 — verified against the archiveBuilt to survive real scrutiny: auditable, deterministic, and honest about the boundary between what the spec says and what it implies.
Every claim is tagged spec-grounded or mechanism-inferred — you always know what you can cite and what is analysis.
Every citation is validated against a registry of 9,535 spec sections and 5,027 tables. Fabricated references cannot survive to display.
Connects behavior across RRC, MAC, and PHY (38.331 / 38.321 / 38.211) to answer the cross-layer questions a single spec section never does.
Unlike conventional RAG, there is no embedding model and no semantic similarity search. Knowledge is compiled into structure once, then retrieved deterministically.
Raw 3GPP specs are compiled by a two-agent pipeline into 511 discrete, structured knowledge concepts — not chunked text.
Answers are grounded on retrieved concepts, labeled grounded / inferred, then every citation is checked against the section registry.
Knowledge gaps trigger targeted ingestion through gated, reviewed updates — the corpus grows without drift.
On a blind SME benchmark of expert 5G NR questions, OKF v2 scored 91.7 / 100 — 24 points ahead of vanilla vector RAG. The same query model and grounding prompt were used for every system; only the corpus and retriever differ.
Cold-start SME eval set v1 · 20 expert questions · blind rubric scoring · comparative signal across systems.
Structuring knowledge at ingestion means each query carries a small, bounded context — and the retrieval layer is commodity infrastructure. The result is deterministic, auditable, and cheap to run.
A ~45,000-token vector-RAG query collapses to a capped ~12 KB structured window (~3,000 tokens) — the platform's own token metric reports up to ~93% lower cost per query, a ~15× reduction.
No embedding model to run, no vector database to host, scale, or re-index. Retrieval is SQLite FTS5 BM25 — commodity, deterministic, near-zero standing infrastructure.
Ingestion runs on a cheaper model tier than live queries, and specs are compiled once — not re-embedded on every question. Cost scales with knowledge, not with traffic.
Cost figures are per-query context vs a 45,000-token traditional-RAG baseline, from the platform's built-in token-efficiency metric. Actual spend varies with question mix and model provider.
Send us a real 3GPP question your team has argued about. We'll show you the grounded answer, the labels, and the verified citation — live.
✉ support@ask3gpp.comAnswers are decision-support — not a substitute for the official 3GPP specifications. Coverage expands via targeted ingestion.