Corpus health

Vector index

Collection
bnl_chunks
Dimensions
768 (nomic-embed-text)
Similarity
Cosine
RAG strategy
Hybrid vector + keyword
Reciprocal rank fusion
Private boost
1.5×
Temporal mode
legal_conservative
Chunk target
600 words · 75 overlap
Vector DB
Qdrant on Colin Docker
10.0.2.10:6333
Hybrid search
Azure AI Search
bnl-legal-search
West Europe · Basic SKU
— Indexed passages
— Source documents
— Active scrapers
— Last ingested

Coverage

Legal categories

—

Family Law

Barneloven, child custody (foreldreansvar), samvær, mediation (mekling), separation and divorce proceedings.

—

Child Welfare

Barnevernloven, omsorgsovertakelse, emergency care orders, foster placement, CPS (barnevernet) case law.

—

Labour Law

Arbeidsmiljøloven, collective agreements (tariffavtaler), Arbeidsretten rulings, dismissal, sick leave obligations.

—

Social Welfare

NAV guidance on sykepenger, dagpenger, AAP, uføretrygd, alderspensjon, yrkesskade and social assistance.

—

Tax Law

Skatteetaten's Skatte-ABC, binding advance rulings (BFU), Skatteklagenemnda decisions, income and capital tax.

—

Administrative Law

Sivilombudet reports, Forvaltningsloven, procedural rights, official complaints, Stortinget oversight.

—

Consumer & Housing

HTU (rental disputes), Finansklagenemnda, Forbrukertilsynet, Forbrukerrådet, Pakkereisenemnda decisions.

—

Immigration & International

UNE (Utlendingsnemnda) decisions, ECHR Art. 8 family rights, EMD case law, Hague Convention (cross-border child abduction).

—

Government Documents

NOUer, Stortingsmeldinger, government white papers and regulatory guidance from Regjeringen.no.

Data sources

Active scrapers

Source Type Category Lang Schedule Status
Loading sources…

Software

AI stack

Reasoning LLMs

  • Azure gpt-4o-mini ★ default — fast, cost-efficient
  • Azure gpt-4o — highest quality
  • GPU qwen2.5:14b — local, private
  • GPU qwen3:14b — reasoning mode
  • GPU dbn-legal-agent — Norwegian law fine-tune (QLoRA on qwen2.5:7b, NorwAI-24B distillation)

All routed via LiteLLM on Colin · 10.0.1.10:4000

Transcription

  • GPU Whisper large-v3 ★ primary
    Cuttlefish · RTX 3060 12 GB VRAM
  • API OpenAI Whisper API
  • Azure AI Speech nb-NO (Norway East)

Speaker diarization · VAD silence filter · beam size 5 · vocabulary presets (barnerett, mediation)

Embeddings

  • nomic-embed-text — 768-dim dense vectors
  • Ollama on Chloe 10.0.1.11:11434
  • Cosine similarity in Qdrant

All documents chunked and embedded before indexing; chunks stored in both Qdrant (vector) and MariaDB (keyword fallback)

Vector & Hybrid Search

  • Qdrant bnl_chunks · ~220 K vectors
    Colin Docker · 10.0.2.10:6333
  • Azure AI Search bnl-legal-search
    Basic SKU · West Europe · hybrid keyword + semantic
  • Reciprocal rank fusion (vector + keyword)
  • Private corpus boosted 1.5×

Chunking pipeline

  • Heading-aware semantic splitting
  • 600-word target · 75-word overlap
  • 50-word minimum chunk
  • SHA-256 deduplication
  • PDF, DOCX, HTML text extraction
  • Temporal metadata (valid_from / valid_until)

Legal temporal reranking: legal_conservative — surfaces current versions first

How it works

Ingestion pipeline

🌐 Source gov websites, APIs, PDFs
🕷 Scraper HTTP / API / PDF
📝 Text extract PDF, DOCX, HTML
✂ TextChunker 600w · 75w overlap
🔢 Embed nomic · 768-dim
⚡ Qdrant cosine upsert
🤖 LiteLLM RAG + LLM
🔍 Your tool Ask, Search, Research…