Built from a chat, line by line.

Bitemporal graph on Neo4j every edge with its own validity window.

Schema from SAT-Graph RAG (JURIX 2025). The whole ontology in a graph database.

• SAT-Graph RAG · arXiv:2505.00039
• LRMoo · IFLA

Two days into the ontology, still zero answers. We kept the bitemporal idea and rebuilt it in SQLite as version_chain (see RAGTAG #08).

BGE-M3 hybrid retrieval dense + sparse + ColBERT, fused at k=60.

BAAI's multilingual model with three signal heads. RRF for fusion, HyDE for English to Finnish query expansion.

• BGE-M3 · BAAI
• ColBERT · SIGIR 2020
• RRF · SIGIR 2009
• HyDE · Gao et al. 2022

Voyage voyage-3-large already ranked the right chunk in the top 30 on every eval question. Adding a second stack failed the cost test.

Courtroom-style debate prosecutor, defense, judge.

Three LLMs arguing per conflict, pattern from AgenticSimLaw.

• AgenticSimLaw · arXiv:2601.21936
• Multi-Agent Debate · NeurIPS 2023

Chat #03 handed us the rule directly. One integer compare resolves every conflict in our eval set; seven LLM turns cost real money.

63,660-node constellation WebGL force layout of the whole corpus.

Attribution view inspired by Anthropic Circuit Tracer.

• @cosmos.gl/graph
• Anthropic Circuit Tracing · 2025

Judges need one answer's reasoning, not the corpus shape. The reasoning panel animates the 5 to 10 nodes that actually mattered.

Live SPARQL fallback CRAG escalation into Semantic Finlex.

Self-RAG reflection tokens to enforce citation coverage at draft time.

• CRAG · arXiv:2401.15884
• Self-RAG · ICLR 2024
• data.finlex.fi/sparql

Cold SPARQL hit 9 seconds on the public endpoint. We can surface 'unsure' offline via AmendmentCaveat instead (RAGTAG #08).

EU-lex contradictions primacy of EU law over national act.

Ingest EUR-Lex directives, model a transposes edge, surface EU vs national conflicts.

• EUR-Lex · hierarchy of norms
• Lex superior · UN OLA

Chat #02 declared this out of scope. Adding it later is not an architectural rewrite: the `transposes` edge type is already in the EdgeType enum, the authority-rank lattice extends to an EU tier with one number, and the strategy router treats it as another `cross_source` route. New corpus, not new infrastructure.

Three deterministic extraction passes structural, anchor, regex.

Edges are emitted by three rule-based passes over HTML and the document tree: structural (parent_of from the heading hierarchy), anchor (cross-references inside an <a href> attribute), and regex (text citations like '§ 102 AVL', 'KHO 2025:46'). No model in the batch graph build.

• src/extraction/structural_edges.py
• src/extraction/anchor_edges.py
• src/extraction/citations_regex.py
• src/extraction/definition_edges.py

Two SQLite tables plus a local LanceDB nodes, edges, chunks. Joined by section_id.

1,967,776 nodes and 2,180,769 typed edges live in two SQLite tables (nodes, edges). 402,088 embedded chunks live in LanceDB on the filesystem (no remote service). Everything joins on section_id with an O(1) lookup.

• scripts/load_graph.py (nodes, edges CREATE TABLE)
• findings/04a_index_sanity.md (402,088 chunks)
• findings/04b_load_report.md (1.97M nodes, 2.18M edges)
• src/indexing/vector_store.py (LanceDB)

Section-anchored chunking 800 to 1,500 tokens, 2,000 hard max, never split mid-citation.

The chunk unit is the SECTION (§). Children are greedily packed under their § head and never split across sentence, item, or citation boundaries. The result: every chunk carries its own legal anchor.

• pipeline/chunks.py

Multilingual embeddings via Voyage voyage-3-large, 1,024-dim, asymmetric query / document.

Hosted but cheap. Asymmetric (input_type='query' vs 'document') to avoid the quality cliff Voyage warns about. Same embedding space carries Finnish, Swedish, and English.

• src/indexing/voyage_client.py (MODEL = voyage-3-large)
• voyageai.com

Strategy router, six presets case_law, recency, definition, cross_source, multi_hop, default.

A keyword and regex classifier on the question text picks one ExpansionStrategy. Each preset sets seed depth, edge types, BFS direction, max hops, and per-edge degree caps. Default falls back to vector-only retrieval.

• src/retrieval/strategy.py

Bounded BFS with hub-skip interprets_in > 30, cites_out > 15, parent_of_in > 50.

Default max_hops = 1. Hub nodes (widely cited statutes) are not expanded through. Final candidate set is truncated to fit a 25k-token context.

• src/retrieval/graph_expand.py

Two reranking paths, one cross-encoder v2 uses bge-reranker-v2-m3, v1 uses metadata signals.

v2 runs BAAI/bge-reranker-v2-m3 (a multilingual cross-encoder) over 30 to 40 candidates and combines 0.6 cross-encoder + 0.3 cosine + 0.1 metadata. v1 (the default path in the API sidecar today) uses a metadata reranker: authority_rank, recency, term overlap.

• src/retrieval/cross_encoder_rerank.py (bge-reranker-v2-m3)
• src/retrieval/rerank.py (metadata reranker)

Temporal correctness version_chain, as_of, AmendmentCaveat. All deterministic.

Every SECTION carries a chronological version_chain of muutetaan, kumotaan, lisätään steps. GraphStore.text_at(section_id, as_of) plays it forward. Every cited chunk on a stale ancestor emits an AmendmentCaveat (suspect, stale, repealed). A separate check_temporal_mismatches function compares the drafted answer against the section's chain via difflib; no LLM in this check.

• src/indexing/graph_store.py (text_at)
• src/models.py (VersionStep, AmendmentCaveat)
• src/agents/verifier.py (check_temporal_mismatches)
• src/retrieval/pipeline_v2.py (wires both)

Authority is one integer Finlex 100, Treaty 90, KHO 80, Vero 60.

Ranks are assigned at ingestion from source / source_subcorpus and stored on every node. Conflict surfacing compares the integer; the team's lattice (Finlex over Vero, KHO can overwrite Vero) drops out of this directly.

• src/extraction/authority.py
• findings/03_authority_ranks.md
• Lex superior · UN OLA

Generation via DeepSeek-V4-Flash hosted on Featherless. Query-rewrite is cached.

The drafter is deepseek-ai/DeepSeek-V4-Flash served via Featherless. Per-question query rewrites are cached in process, which lowers cost on repeated framings. Per-query answers are not cached today; a localStorage history in the Next.js UI lets the user recall past questions but does not skip the call.

• src/retrieval/generate.py (MODEL = deepseek-ai/DeepSeek-V4-Flash)
• src/retrieval/query_rewrite.py (in-process cache)
• Featherless · featherless.ai

• SAT-Graph RAGde Martim, JURIX 2025 · arXiv:2505.00039
• TG-RAGHan et al., 2025 · arXiv:2510.13590
• LRMoo v1.1.1IFLA, 2026 · cidoc-crm.org/LRMoo
• Semantic FinlexSeCo Aalto + MoJ · data.finlex.fi/sparql
• Self-RAGAsai et al., 2024 · ICLR 2024
• CRAGYan et al., 2024 · arXiv:2401.15884
• HyDEGao et al., 2022 · Precise Zero-Shot Dense Retrieval
• AgenticSimLawJan 2026 · arXiv:2601.21936
• Multi-Agent DebateDu et al., NeurIPS 2023 · arXiv:2305.14325
• BGE-M3Chen et al., BAAI · bge-model.com
• bge-reranker-v2-m3BAAI · bge-model.com
• ColBERTKhattab + Zaharia · SIGIR 2020
• RRFCormack et al. · SIGIR 2009
• Voyage voyage-3-largeVoyage AI · voyageai.com
• Anthropic Circuit TracingAnthropic, 2025 · transformer-circuits.pub
• UN OLA · lex superiorUN Office of Legal Affairs · a_cn4_l682.pdf