TIER S BLUEPRINT

Lithuanian Document Q&A (RAG)

Ask your own documents a question in Lithuanian and get an answer with a citation behind every sentence - or an honest "nežinau". LT retrieval quality generic chatbots don't reach.

from €1,990

LT RAG - KLAUSIMAS ↔ ŠALTINIS · LIVE

Kokia yra sutarties nutraukimo tvarka?

sutartis → sutarties · sutartyje · sutartį §7.2 · 0.91 §9.1 · 0.44 §3.4 · 0.21

Nutraukti gali bet kuri Šalis, raštu informavusi kitą prieš 30 kalendorinių dienų[1] §7.2

grounded ✓ 1/1 sakiniai

Kiek kainuoja sutarties pratęsimas?

§3.4 · 0.31 §7.2 · 0.19 žemiau slenksčio 0.55

„Nežinau - to jūsų dokumentuose nėra.“

→ HUMAN · nieko neišgalvota

PASLAUGŲ TEIKIMO SUTARTIS Nr. 2026/114

2026-03-02 · Vilnius

3.4. Kaina ir atsiskaitymas

Užsakovas moka 1 200,00 € per mėnesį. Sąskaita pateikiama iki mėnesio 5 d.

7.2. Sutarties nutraukimas

Sutartis gali būti nutraukta bet kurios Šalies iniciatyva, apie tai raštu informavus kitą Šalį prieš 30 (trisdešimt) kalendorinių dienų.

9.1. Ginčų sprendimas

Ginčai sprendžiami derybomis, nepavykus - Vilniaus komerciniame arbitraže.

psl. 3 iš 7 · indeksuota 2026-07-14 · eu-central-1

audit_log ▸ #7734 sha256:4b1a…c7 ⛓ chain verified · groundedness 1.00 · abstain 1/2 · region: eu-central-1

How it works

Anyone can demo a chatbot on one clean PDF. The build starts where the demo ends: messy ingestion, an inflected language, and knowing when to stay quiet.

Ingestion is the unglamorous half

Point it at SharePoint, S3 or a mailbox. PDFs, scans (OCR), tables and DOCX get parsed and split on Lithuanian sentence boundaries - not every 512 characters, which is how you end up retrieving half a clause. Documents re-index when they change. Nothing leaves the EU.

Lithuanian retrieval, not translated English

"Sutartis", "sutarties", "sutartyje" are one word to a reader and three to a generic embedder. Hybrid retrieval - LT-tuned embeddings plus BM25 over lemmatized text - finds the passage that English-first RAG walks straight past.

No citation, no sentence

Every sentence in the answer points at the exact passage it came from. Click it and the source page highlights. A claim with no passage behind it doesn't get written at all.

It says "nežinau"

When retrieval comes back weak, the assistant abstains and routes to a human instead of writing something plausible. The abstention rate is a tuned, published number - not something you discover in month two.

The Four Guarantees™ - this build

Measured value

Eval gate before ship, on an eval set built from your own questions: retrieval recall@k, groundedness (does every sentence trace to a retrieved passage), and abstention calibration - does "nežinau" show up exactly when it should.

Defensible

Click-to-source on every sentence. Retrieved chunks, model version, prompt and answer logged per query, in a tamper-evident hash-chained audit log. Annex IV documentation regenerated every run.

Self-correcting

Every thumbs-down and every human correction becomes a case in the eval set. Retrieval drift alerts when the corpus shifts under the index. The eval set grows into the thing that protects you.

Yours & everywhere

One container, your cloud, full source handover. MCP tools so your own agents can query the corpus directly - with the same citations and the same right to abstain.

The claims, sized honestly

Reference buyer: LT/Baltic firm with a few thousand internal documents - contracts, policies, technical manuals - and a team answering the same questions over and over.

100% of answer sentences carry a click-to-source citation

0 answers invented when retrieval comes up empty - it abstains

LT lemma-aware hybrid retrieval, not English embeddings pointed at Lithuanian

EU ingestion, index and inference - documents never leave the region

Three ways to own it

Tier	What you get	Price
Scaffolding ★	The full repo - ingestion pipeline, LT hybrid retrieval, citation-bound answering, abstention gate, eval harness, audit log, MCP server. Runs on a public LT corpus so you can watch it work before your data touches it. You own the source.	€1,990
PoC ★★RECOMMENDED	Pointed at your real documents. Retrieval recall, groundedness and abstention calibration measured against an eval set built from your own questions. Full code and the quality report are yours. No accuracy guarantee at this stage, by design.	from €6,000
Implementation ★★★	Production: wired into your document sources, permissions-aware retrieval, human routing live, monitoring + drift alerts, re-indexing as documents change, monthly groundedness and abstention report - the agreed number (conservative) attaches here.	from €18,000

★ = engagement depth. PoC is the recommended path: quality proven on your data before production money. The PoC carries no performance guarantee by design; the agreed number (conservative) attaches at Implementation, informed by the PoC report.

What we don't promise

It will not answer everything, and that's the part you're paying for. Questions your corpus doesn't actually cover come back as "nežinau" with a pointer to the closest documents - by design. An assistant that always has an answer is one that sometimes invents it, and you find out which in front of a client. We publish the abstention rate monthly. If it climbs, that's a hole in your documentation, not a model that got worse.

Ready to see your own number?

Request the build: within 48h you get a personal reply with the value sized to your volume.

No commitment · reply within 48h · your data stays in the EU