LatestContext is everything in the age of legal AIRead the article
Engineering · 3 min read

Grounding, not guessing: the retrieval stack behind every PONS answer

Parametric memory cannot cite, cannot date itself, and blends jurisdictions. Here is the actual machinery PONS runs instead - hybrid search, query classification, authority-weighted ranking, and token-budgeted context assembly.

Sebastian Melbye · 17 March 2026

Why a model's memory fails legal work

Grounding is a simple promise to the firm: every answer PONS gives is built from real documents, retrieved for that question, with the sources shown. Not a confident paragraph from a model's memory - an answer you can rely on, check, and stand behind with a client, with no invented citations.

A general model answering from memory has three defects legal work cannot tolerate. Its memory is undated: it cannot tell whether a rule still held on the date that matters. It blends jurisdictions: Norwegian, German, and English law merged into the same weights. And pressed for a source, it generates something citation-shaped rather than retrieving something real. So PONS runs on one contract: the model never answers from memory. This post is the machinery that enforces it.

The retrieval stack

Every query goes through three layers. Keyword matching handles exact statute numbers and terms of art. Vector search catches conceptual matches sharing no vocabulary with the query. A semantic reranker then orders the results by meaning, not raw text overlap.

Answers come from documents, not memory ?Your questionThe documents§The answer, cited

Index topology matters as much as the algorithm. Each firm's material lives in its own per-organization indexes - one organization's index is never a candidate for another's query. The curated legal corpus is organized by jurisdiction: a Swedish case-law question searches Swedish case law, not a pan-European soup. A score floor keeps weak matches out of the prompt - fewer strong passages beat one more that is noise.

Retrieval that knows what kind of question it got

Legal queries come in two shapes wanting opposite strategies. "SFS 2018:218" is a citation - the user wants exact matching, nothing thematic. "When can an employer terminate for redundancy?" is conceptual - meaning matters, vocabulary barely does. Every query is classified as citation, conceptual, or mixed, and the balance between exact and semantic search shifts to match. Smaller refinements - expanding legal acronyms to full statute names, stopping one long statute from filling the result list - do quiet work.

Ranking beyond similarity

Two passages can be equally similar to a query while one is a current supreme court decision and the other a repealed district ruling. So the corpus indexes add two signals beyond similarity. Freshness is keyed to an instrument's effective date, so current law rises and stale law decays. Authority weighting encodes the hierarchy of legal sources - constitution and treaty above supreme court, supreme court above legislation, legislation above appellate rulings.

The ranking function knows what the sources are, not just what they say - which is exactly what a lawyer would expect from a junior associate.

Ranked by more than similarity Authority · recency · similarityThe right source first

Receipts, not decoration

Retrieval finds more good material than a prompt holds, so context is assembled under an explicit token budget - the strongest material in full, the rest summarized or dropped.

Grounding only counts if you can audit it. Every snippet carries exact character positions back into its source document, so the interface can highlight the precise passage behind a claim. Sources are numbered, cited inline, and only the ones the model actually used are shown.

The citation list is not a bibliography of what we searched. It is a receipt for what the model used.

Retrieval draws from the curated legal corpus, the firm's knowledge base, and the matter's own documents. Firm and matter material stays in per-organization indexes, customer data is never used to train models, and everything stays within EU data residency - isolation enforced by index topology, not by a policy document.

The honest limit

None of this makes the model smarter. It makes the model honest about its evidence, so answer quality is bounded by retrieval quality, and retrieval by ingestion quality. A scanned PDF whose text never extracted, a clause split awkwardly across chunks - each is a passage retrieval cannot find, and a missing passage is a claim the model cannot make. That is why so much engineering effort sits upstream of the model, in parsing, structure reconstruction, and corpus curation.

Not a model that knows the law. A system that finds the law, answers from it, and shows you its receipts.

See it on your own documents

Legal work,delivered with PONS.