lex-au-search-index

Pre-built Qdrant vector index for hybrid dense + BM25 search over Australian Commonwealth legislation - download and query, no re-embedding or re-ingest required.

This is an output of lex-au-search run against the lex-au corpus. If you want to search AU legislation, download this and point the CLI at it.

Version: v0.7.0, index generation 3 - 553,870 chunks (sections, subsections, schedule clauses, schedule text) across 3,086 Acts, embedded with snowflake/snowflake-arctic-embed-l (1024-dim, local ONNX, 512-token context). Corpus pinned at lex-au commit ce532dde (lex-au v0.13.0 content, 2026-10-26). Built from embed cache revision c8a24dfc. Earlier generations stay reachable by revision: generation 2 (552,305 chunks, 3,085 Acts) is 461bc8aa4ae8bbbe0f0a28fcc5485f1376451f21.

See github.com/cchew/lex-au-search for source code, CLI, and MCP server. See cchew/lex-au-search-embed-cache for the separate raw embedding cache the ingest pipeline itself uses to resume/skip re-embedding.

Quick example

Download the index and run the search server directly - no embedding model needed to browse:

from huggingface_hub import snapshot_download
from qdrant_client import QdrantClient

path = snapshot_download(repo_id="cchew/lex-au-search-index", repo_type="dataset")
client = QdrantClient(path=f"{path}/qdrant_storage")

# Browse Act-level metadata with no query embedding required
hits, _ = client.scroll("legislation", limit=3, with_payload=True)
for h in hits:
    print(h.payload["act_name"], h.payload["frbr_uri"])

For real hybrid dense + BM25 search (search_legislation), clone lex-au-search, pip install -e ".[dev]", then:

lex-au-search serve --storage-dir /path/to/downloaded/qdrant_storage
# or: LEXAU_SEARCH_STORAGE=/path/to/downloaded/qdrant_storage lex-au-search mcp

What's in this dataset

A Qdrant local-mode storage directory (qdrant_storage/) with two collections:

• legislation (3,086 points) - one per Act: act_name, frbr_uri, year, as_at_date, section_count, schedule_clause_count.
• legislation_section (553,870 points) - one per chunk: 1024-dim dense vector (snowflake-arctic-embed-l) + BM25 sparse vector, payload act_name, frbr_uri, eid, provision_num, provision_type, schedule_num (the schedule's printed number, e.g. 3A, on schedule chunks), heading, text (full chunk text), refs (cross-references).

On-disk size ~6.4 GB, dominated by legislation_section (vectors + full section text payload).

Known limits

• Dense embeddings see only a chunk's first 512 tokens, so very long chunks (about 1,200 exceed 2,000 words) rank poorly even for verbatim queries.
• Dictionary Schedules (e.g. Criminal Code Schedule 1) are cut into 400-word windows, not one chunk per definition, so a single definition can rank low and cannot be cited on its own.
• Schedule clauses that repeat the same criterion text across visa subclasses (Migration Regulations Schedule 2) are hard to tell apart unless the query names the subclass.
• Qdrant's local/embedded mode keeps the full vector set resident in RAM and has no ANN index, so QdrantClient(path=...) construction itself is slow at this scale (multiple minutes to open, 550K+ points) - expected, not a bug. Fine for local CLI/MCP use; not suited to serving concurrent requests.
• No hosted search endpoint exists for this index - it's a downloadable artifact, not a live service.

Licence

CC BY 4.0. Source legislation is Crown copyright - Commonwealth of Australia; reproduction permitted for non-commercial and research purposes under the PSI Framework. The legislation_section payload carries substantial verbatim section text from that corpus, so this dataset inherits the same licence as cchew/lex-au rather than a separate one for "just the vectors."

What's built on this

• ClauseKit - LLM extraction of evaluatable rules (JSON Logic) from Acts, run claims against them grounded back to source clauses
• Act Alike - compare how terms are defined across Acts

Related

• github.com/cchew/lex-au-search - source code, CLI, MCP server
• cchew/lex-au - the underlying AKN 3.0 XML corpus this index was built from
• cchew/lex-au-search-embed-cache - the ingest pipeline's own resumable embed cache (different artifact, different audience)
• lex-au-graph - cross-reference graph and definition resolution across Acts