transport-lit — transportation grey literature over MCP

(Renamed from dot-lit on 2026-08-27; DOT_LIT_* environment variables and the old data directory are still recognised.)

transport-lit gives an AI assistant (Claude Desktop, Claude Code, any MCP client) keyword search over the transportation research reports that PubMed does not index and Semantic Scholar covers poorly. It started with ROSA-P, the U.S. National Transportation Library's repository (NHTSA DOT HS reports, FHWA/FRA/FTA/FAA, UTC and state DOT research; https://rosap.ntl.bts.gov), and now harvests six OAI-PMH sources on three continents plus whatever you export from TRID:

transport-lit sources lists them; transport-lit harvest --source <key>|all harvests them; the collection filter in search_reports selects one (e.g. "VTI", "BASt", "World Bank", "CEPAL", "TRID"). Adding another OAI-PMH repository is one entry in src/transport_lit/sources.py.

I created this for my own personal academic and research use and am happy to share it with anyone else who finds it useful. I welcome feedback on errors, integration needs, improvements and other commentaries. I will check those regularly and will integrate them as much as possible and document that. If you are interested in helping to support this or have other ideas for it, I welcome them! — Alex Quistberg (open an issue)

It does this the only way that works for an OAI-PMH source: it harvests the whole repository's metadata into a local SQLite database, builds an FTS5 full-text index over it, and serves search from that index. Nothing is queried live except an optional PDF fetch for full text. Re-harvests are incremental (from= on the OAI request) and cheap.

Tool surface

Plus one prompt, literature_scan(topic), that walks a model through a multi-query scan with citations. Every tool carries MCP annotations (readOnlyHint, idempotentHint; only get_fulltext is openWorldHint because it may fetch one PDF).

id accepts dot:93144, 93144, oai:dot.stacks:dot:93144, or the landing URL. Imported records use other prefixes (trid:813520, import:…).

Query syntax: bare words are ANDed first; if fewer than limit hits match every term the remaining slots are filled with any-term matches (match_mode = all_terms / any_terms). Quote phrases ("driver improvement"), use a trailing * for a prefix. Ranking is BM25 with title, report number and author weighted above abstract.

Related MCP server: Personal Research Assistant MCP

Setup

Requires Python 3.12+ and uv.

git clone https://github.com/aquistbe/transport-lit && cd transport-lit
uv tool install .            # installs `transport-lit` (CLI) and `transport-lit-mcp` (server) on PATH
export TRANSPORT_LIT_CONTACT=you@example.org   # identifies your harvester to ROSA-P (put it in your shell profile)
transport-lit probe                # live check: Identify / ListMetadataFormats / ListSets
transport-lit harvest              # full harvest the first time (~15 min), incremental afterwards
transport-lit status               # counts, last run, coverage by year
transport-lit search driver improvement program evaluation

For development use uv sync and prefix commands with uv run (e.g. uv run pytest).

Any MCP client, any model

The server speaks standard MCP over stdio (default) and Streamable HTTP / SSE (transport-lit-mcp --transport streamable-http --port 8765, endpoint /mcp). transport-lit mcp-config [client] prints a ready-to-paste snippet for: Claude Desktop, Claude Code, Cursor, VS Code (Copilot agent mode), Zed, Continue, LM Studio, Goose, Open WebUI and LibreChat (the last two over HTTP). A Dockerfile builds an HTTP server image with the index on a volume.

Open models. Tested end to end on 2026-08-26 with Ollama qwen2.5:3b (3 B parameters) via tests/ollama_smoke.py: given "find reports about driver improvement programs; list 3 titles with years and ids", the model called search_reports({"query": "driver improvement", "limit": 3}) once and answered with correct titles, years, ids and landing URLs from three sources. Design choices that make small models work: ten tools with one-line-first descriptions, flat JSON arguments with defaults, compact hit objects (no raw metadata in search results), and a server instructions string that names the sources and filters. Run the smoke test with any tool-capable model: OLLAMA_MODEL=llama3.1 uv run python tests/ollama_smoke.py "…".

Register in Claude Desktop

transport-lit install-claude-desktop          # prints the JSON to add
transport-lit install-claude-desktop --write  # merges it into claude_desktop_config.json (keeps a .bak)

The entry it writes is simply:

{ "mcpServers": { "transport-lit": { "command": "/Users/you/.local/bin/transport-lit-mcp", "args": [],
                               "env": { "TRANSPORT_LIT_DATA_DIR": "/Users/you/.local/share/transport-lit",
                                        "TRANSPORT_LIT_CONTACT": "you@example.org" } } } }

Restart Claude Desktop afterwards. For Claude Code: claude mcp add transport-lit -- transport-lit-mcp.

Configuration (environment variables)

Variables can also live in ~/.config/transport-lit/env as KEY=VALUE lines (real environment variables win); that is where a NII application ID or NCBI key belongs, so the launchd jobs, the MCP server and manual runs all see it. No credentials are required; all endpoints are public.

Install from PyPI (no clone)

uv tool install transport-lit            # CLI + MCP server on PATH
uvx --from transport-lit transport-lit-mcp   # or run the server ad hoc
uv tool install "transport-lit[semantic]"    # with the bundled embedding backend

Published at https://pypi.org/project/transport-lit/ through GitHub's trusted publishing: a tag runs release.yml (build, GitHub release), which then dispatches publish.yml — the one workflow registered as a trusted publisher — where the upload waits for the maintainer's approval on the pypi environment. PyPI rejects tokens from reusable workflows and matches the top-level workflow file, which is why publishing is a separate dispatched workflow rather than a job inside the release. server.json is the manifest for the MCP Registry (registry.modelcontextprotocol.io), to submit after the PyPI package exists.

Pinned versions

Releases are git tags vMAJOR.MINOR.PATCH (semantic versioning: patch = fixes, minor = new tools/sources, major = a breaking change to the tool surface or database schema). Each tag triggers the release workflow, which runs the tests, builds a wheel + sdist and attaches them to a GitHub Release. Python dependencies are pinned by the committed uv.lock; CI installs with uv sync --frozen, so a release always runs against the exact versions it was tested with. To install a specific version:

uv tool install "transport-lit==0.3.0"                            # a pinned PyPI release
uv tool install git+https://github.com/aquistbe/transport-lit@v0.3.0  # or the matching git tag
uv tool upgrade transport-lit                                     # move to the latest release

Harvesting

transport-lit harvest                     # ROSA-P; auto: incremental if a complete full harvest exists, else full
transport-lit harvest --source all        # every configured source (vti, bast, wbokr, ipea, cepal, rosap)
transport-lit harvest --mode full         # walk the whole repository again
transport-lit harvest --mode incremental  # from = start of last complete run − 1 h, until = now
transport-lit harvest --from 2026-08-01T00:00:00Z   # explicit window (full timestamp required)
transport-lit harvest --max-pages 3       # testing only; the run is recorded as failed/partial
transport-lit reindex                     # re-parse the cached raw pages (no network) after a parser change

What the harvester does and why (all behaviour verified against ROSA-P on 2026-08-26):

• ListRecords&metadataPrefix=oai_dc, 100 records per page, following resumptionToken until a page arrives without one. Only then is the run marked complete; any error leaves it failed and does not advance the "last harvest" pointer, so harvest_status never claims a partial index is complete.

• Pacing: one request per TRANSPORT_LIT_MIN_INTERVAL seconds (default 1 s). Tokens expire about 60 s after issue, so retries use short backoff (2/4/6 s).

• badResumptionToken, transport errors, truncated XML, or an empty envelope while a token is live → the list is re-issued. ROSA-P does not return records in a stable datestamp order (checked on every page), so recovery restarts the list from the top; upserts make that idempotent. If ordering had been monotone the harvester would instead resume from the smallest datestamp seen via until=. Up to 8 recoveries per run, then failed.

• noRecordsMatch: ROSA-P does not send the error code; an empty selective harvest comes back as an OAI-PMH envelope with no <ListRecords> element. That is mapped to "nothing to do" only when no token was in play; mid-list it is treated as truncation.

• Silent truncation checks: the token's cursor is compared with the local count on every page; a full harvest that returns >5 % fewer records than the previous full harvest is flagged in the run notes. Both appear in harvest_status().last_harvest.notes.

• Deletions: the repository reports deletedRecord=no, so nothing is ever removed locally; a record that vanishes from ROSA-P stays in the index until a full re-harvest into a fresh TRANSPORT_LIT_DATA_DIR.

• Caching: every OAI page is stored gzipped under raw/run<N>-p<page>.xml.gz, so the parser can be changed and the index rebuilt without touching the network; PDFs and their extracted text are cached under pdf/ and in the fulltext table.

• from/until are formatted to each repository's declared granularity (read from Identify): ROSA-P, DiVA and DSpace take full YYYY-MM-DDThh:mm:ssZ timestamps, OPUS (BASt) takes only YYYY-MM-DD and the window is widened a day each side.

• Broad repositories (World Bank, IPEA, CEPAL) are filtered at harvest time by a multilingual transport vocabulary (sources.TRANSPORT_RE, en/es/pt/de/fr/sv): a record is kept if a term appears in the title, or (IPEA, CEPAL) if two distinct terms appear among the subject headings. Abstracts are ignored — development literature mentions roads and ports in passing — and World Bank subjects are ignored too (100+ headings per record). This was tuned on 2026-08-26 against random 20-title samples: the loose title+subjects+abstract rule kept 15,271 World Bank records at roughly 35–50 % precision; the final rule keeps 976 at 18/20, IPEA 207 at ~16/20, CEPAL 1,165 at ~15/20. Recall is the price; loosen min_subject_hits in sources.py and run transport-lit reindex --source <key> (no network) if you want the other trade. The run notes record kept vs skipped.

• transport-lit doctor [--repair] checks SQLite integrity, both FTS indexes, runs stuck in running, impossible timestamps and WAL size, and repairs what is safe; the store also checkpoints the WAL on close. transport-lit cite prefetch [--source …] resolves every DOI/PMID/OpenAlex-bearing record to its OpenAlex work in bulk (50 per request) so cited_by_count is known for them without a per-record call.

transport-lit reindex --source <key> re-parses the cached pages and prunes records the current parser/filter no longer keeps, so filter changes never need a re-harvest.

Monthly rebuild and weekly updates (maintenance schedule)

The corpus changes slowly, so the cadence is: weekly incremental harvest and a monthly fresh rebuild. harvest --fresh full-harvests into a temporary store and then atomically replaces the dot: records in the live index — the only way records ROSA-P stops serving ever disappear (its OAI-PMH endpoint does not track deletions). Imported sources (TRID exports) are untouched, and a failed rebuild changes nothing.

transport-lit install-schedule          # shows the two launchd agents
transport-lit install-schedule --write  # installs them: Mon 06:00 `--source all` incremental, 1st 05:00 `--source all --fresh`

Logs land in $TRANSPORT_LIT_DATA_DIR/logs/. On Linux use the cron lines the command prints.

Monthly maintenance checklist (done with the rebuild): read new GitHub issues; uv lock --upgrade && uv run pytest; note fixes in the changelog section of the release; bump version in pyproject.toml and src/transport_lit/__init__.py; git tag vX.Y.Z && git push --tags.

TRID: import what you export

TRID (https://trid.trb.org) is the most complete transportation bibliography and the natural complement to ROSA-P, but it has no API, its FAQ says TRB "does not grant access to TRID backend systems or lift export/download restrictions", and its robots.txt disallows AI crawlers. What every user may do is search and export. So:

1. Run your search in TRID, choose Export → RIS (CSV and XML are also offered).

2. transport-lit import ~/Downloads/trid-driver-improvement.ris --collection "TRID: driver improvement"

Records get ids trid:<accession> from the TRID view URL, land in the TRID collection (search_reports(..., collection="TRID")), and re-importing the same file is idempotent. The importer is generic RIS, so Zotero/EndNote/Scopus exports work the same way with --source <prefix>. get_fulltext on an imported record only follows a direct .pdf link; otherwise use landing_url.

What ROSA-P's OAI-PMH endpoint offers

https://rosap.ntl.bts.gov/fedora/oai — repository "DOT Stacks" (the CDC Stacks platform), protocol 2.0, earliest datestamp 2008-07-02, no deletion tracking, no OAI sets (ListSets is empty), and oai_dc is the only metadata format. It is a qualified Dublin Core in disguise, though: elements such as dc:contributor.author, dc:description.abstract, dc:relation.isPartOf, dc:identifier.uri (DOI and report numbers, e.g. DOT HS 813 827), dc:coverage.spatial, dc:title.alternative and dc:description.tableOfContents are all present. The parser (dc.py) keeps every raw field and derives the typed columns from them. dc:relation.isPartOf (semicolon-separated) is what list_collections / the collection filter use.

PDF links are not in the metadata; get_fulltext reads citation_pdf_url from the landing page and falls back to the datastream convention /view/dot/{n}/dot_{n}_DS1.pdf.

Verification (2026-08-26)

v0.4.0 semantic search. 342,462 vectors (fastembed multilingual MiniLM-L12, 384-d, data-parallel at 94 records/s on 8 cores — 60 min for the corpus). Cross-language check: "elderly pedestrian crashes at night" in semantic mode returns, among its top 8, three Japanese-language CiNii reports (夜間 高齢歩行者 死亡事故 analyses, 1995–2011) beside PubMed and English CiNii items; vocabulary check: "point system for problem drivers license suspension recidivism" finds ROSA-P's 1997 California vehicle-impoundment evaluation and 1986 administrative-revocation report, which share no query words. Hybrid latency ≈ 0.6 s (query encoding dominates), keyword ≈ 25 ms. Operational lesson recorded here so nobody repeats it: never delete a SQLite -wal file while another process (e.g. a running MCP server) has the database open — it holds committed data not yet checkpointed.

v0.3.0 API sources. OpenAlex: 58 pages, 11,448 reports (10 topics, type:report), 1,428 with PDF links. CiNii: 730 pages, 144,348 hits over 20 queries, 118,609 unique. PubMed: 105,028 articles in 17 date slices (E-utilities caps retstart at 10,000, so slices are found recursively). Open-model check: Ollama qwen2.5:3b answered a driver-improvement question with one correct search_reports call.

v0.2.0 multi-source harvest. VTI: 120 pages, 11,944 seen, 11,460 unique (DiVA serves some records in several sets), 0 resumptions. BASt: 30 pages, 2,987 seen, 2,970 unique; 1,901 with direct PDF links; day-granularity incremental path exercised (24 records). World Bank: 404 pages / 40,332 seen; IPEA: 144 / 14,400; CEPAL: 522 / 52,199 — all ended on a token-less page with 0 resumptions; filtered counts above. Spot searches: Fußgänger Unfall (BASt) → crash reconstruction and rural-road crash statistics; acidentes de trânsito mortalidade (IPEA) → "Mortalidade por acidentes de transporte terrestre e desigualdades interestaduais no Brasil"; seguridad vial peatones (CEPAL) → road-safety governance and campaign evaluations; pedestrian safety (VTI) → 1990s child-pedestrian training studies.

v0.1.0 (first ROSA-P harvest)

Harvest completeness. Run 1 (full) walked 908 pages / 90,706 records in 15 min (00:03:59–00:19:11 UTC) with 0 resumptions, 0 cursor mismatches, and ended on a page of 6 records with no resumption token — the OAI-PMH definition of a complete list. 90,603 unique records are in the store; the 103-record gap is the same record appearing on two pages, which happens because ROSA-P does not return records in a stable order (the harvester logs this: "datestamp ordering violated on page 2"). A second independent full pass, 30 minutes later into a separate directory, returned exactly the same numbers — 908 pages, 90,706 seen, 90,603 unique — and the two ID sets are identical (0 records unique to either pass). The 103 repeats are the repository serving the same record on two pages, not records being skipped.

Coverage by decade (year present for 74,448 = 82 %; the remaining 16,155 have no date in any metadata field; year_source says whether a year came from dc:date (48,658), a bare-year description line (22,205) or the title (3,585)):

Known-item retrieval (transport-lit search …, rank 1 unless noted):

Real query driver improvement program evaluation negligent operator (top 6 of 10):

1. dot:18905 — An evaluation of the short-term effects of the Virginia driver improvement program (Lynn) — all_terms

2. dot:29326 — Review of NJ point system (Carnegie, Ozbay, Mudigonda, 2013; FHWA NJ-2013-004) — all_terms

3. dot:18959 — …Virginia driver improvement program on negligent driving: 12-month report (Lynn)

4. dot:18969 — …Virginia driver improvement program on negligent driving: 24-month report (Lynn)

5. dot:17678 — Study of recidivism rates among drivers administratively sanctioned by the New Jersey MVC (Carnegie et al., 2009)

6. dot:17677 — Study of the effects of plea bargaining motor vehicle offenses (Carnegie et al., 2009)

Full-text extraction was checked on dot:93144 (DOT HS 813 827, 3.7 MB PDF, resolved via citation_pdf_url). Unit tests: uv run pytest (parser for both metadata profiles, year fallback, FTS search/filters, upsert idempotence, query tokenizer, id normalisation).

Layout

src/transport_lit/
  config.py    paths, User-Agent, pacing, limits (env-overridable)
  oai.py       rate-limited OAI-PMH client; typed errors; raw-page cache
  dc.py        oai_dc record -> typed dict (authors, year, DOI, report numbers, collections …)
  store.py     SQLite schema, FTS5 index + triggers, search, stats, harvest-run bookkeeping
  harvest.py   full / incremental harvest with completeness + truncation handling
  fulltext.py  PDF resolution, download (size-capped), pypdf extraction, cache
  server.py    MCP tools (FastMCP / MCPServer)
  importers.py RIS import (TRID exports and any other reference-manager export)
  cli.py       transport-lit probe | harvest [--fresh] | import | reindex | status | search | get | fulltext
               | install-claude-desktop | install-schedule
.github/workflows/  ci.yml (tests on push/PR), release.yml (wheel + GitHub Release on tag)
tests/         unit tests (parser, store, query tokenizer)

Adding a second source later (e.g. NHTSA crashstats)

The store is source-agnostic: records.id is a prefixed string (dot:93144 today), harvest_runs.source records which harvester wrote a run, and the FTS index does not care where a row came from. To add a source:

1. Write src/transport_lit/sources/<name>.py exposing harvest(store, *, mode, progress) that yields dicts in the same shape dc.parse_record produces (id, title, authors, year, abstract, report_numbers, doi, landing_url, collections, raw, …) and calls store.upsert_records(). Use a new id prefix (nhtsa:812115) and pass your own source name to store.start_run() so harvest_status can report it separately.

2. Reuse oai.RateLimiter and config.USER_AGENT for etiquette; store raw responses under raw/<source>/ for reproducibility.

3. Give harvest.status() a per-source block (count by id prefix).

4. Add a --source option to transport-lit harvest and, if the source has its own facet, a corresponding filter on search_reports.

5. Dedupe against ROSA-P by DOI / report number (records.doi, records.report_numbers) rather than by title — NHTSA reports are often present in both places.

Verified facts for the NHTSA crashstats source, so nobody re-derives them: https://crashstats.nhtsa.dot.gov/Api/Public/Publication/{id} returns the PDF directly (812115 → NMVCCS critical-reasons report, application/pdf, ~0.5 MB). It is a document-retrieval endpoint, not a search or listing API, so a connector will need an enumeration strategy (e.g. the DOT HS numbers already present in ROSA-P report_numbers) rather than a crawl.

Beyond the U.S.: candidate sources assessed (2026-08-26)

Probed live for (a) whether the holdings are literature rather than datasets and (b) whether there is machine access that fits this harvester. Counts are what the endpoints reported that day.

By region (same-day probes; "open" means unauthenticated machine access confirmed):

The three OAI-PMH repositories with completeListSize (VTI, BASt, World Bank OKR) fit the existing harvester with a source prefix and a per-source metadataPrefix; DSpace 7 sites also tolerate from/until and return proper noRecordsMatch, so the ROSA-P quirks in oai.py are already the harder case.

TRID is out of scope

TRID (https://trid.trb.org) has no public API, no OAI-PMH endpoint and no bulk export. Its FAQ states that "TRB does not grant access to TRID backend systems or lift export/download restrictions for individuals or organizations" and that the database may not be used to train LLMs. It is deliberately not scraped here.

v2 order (agreed 2026-08-26)

1. VTI + BASt (done, v0.2.0) — 2. World Bank OKR, IPEA, CEPAL (done, v0.2.0) —

2. IRDB Japan — 4. OpenAlex type:report as global backstop — 5. a PubMed transport subset (see below). VTI note: DiVA's oai_dc carries no full-text link; switching that source to swepub_mods/mets_kb would give get_fulltext the FULLTEXT01.pdf URL.

PubMed: a transport/injury subset, not all of PubMed

PubMed's E-utilities (esearch/efetch, free, 3 req/s without a key) can maintain a local subset from a fixed strategy, refreshed with mindate/maxdate on the same weekly/monthly cadence. Two complementary filters, OR-ed together:

• MeSH strategy — "Accidents, Traffic"[MeSH] OR "Pedestrians"[MeSH] OR "Bicycling"[MeSH] OR "Automobile Driving"[MeSH] OR "Motorcycles"[MeSH] OR "Wounds and Injuries"[MeSH] AND ("Transportation"[MeSH] OR "Built Environment"[MeSH] OR "City Planning"[MeSH]) — catches transport papers in general and clinical journals.

• Journal list — Accident Analysis & Prevention, Traffic Injury Prevention, Journal of Safety Research, Injury Prevention, Injury Epidemiology, Journal of Transport & Health, Safety Science, Transportation Research Part F, Transport Reviews, BMC Public Health (transport-tagged only), etc. — catches transport papers indexed without the MeSH terms.

SafetyLit (safetylit.org, the WHO-affiliated weekly injury-literature bulletin) maintains exactly such a journal list and hand-classifies articles by topic, which would make it the best seed for the journal filter; its site was unreachable (connection refused on every host name) when checked on 2026-08-26, so its current status is unconfirmed.

Weekly digest (a SafetyLit-style bulletin)

transport-lit digest --days 7 [--abstracts] prints a Markdown bulletin of everything that entered the index in the last week, grouped by source, with counts. It is driven by first_seen_at, which is set the first time a record is seen and preserved across fresh rebuilds, so a monthly rebuild does not make the whole index look new. The whats_new tool exposes the same data to a model, which can then write the summaries — the editorial step SafetyLit did by hand.

Compared with other literature MCPs

PubMed, Semantic Scholar, OpenAlex and arXiv MCP servers proxy live queries to one API. transport-lit differs in three ways: it indexes grey literature the aggregators lack (agency reports, state DOT evaluations, ITRD-contributing institutes), it runs offline on a local index after harvesting (no rate limits at query time, no key), and it is multi-source with one id scheme, so a model can search everything at once and export citations. What those servers have that this one still lacks: citation graphs (who cites whom), author disambiguation, and semantic (embedding) search — see below.

Citation graph (v0.5)

get_references / get_citations (CLI: transport-lit cite refs|cites <id> [--in-index]) attach OpenAlex's citation graph to the index. A record is matched to an OpenAlex work by its OpenAlex id, DOI, PMID, or — for the many undated, DOI-less agency reports — an exact normalised title with the year within ±1 (match in the result says which). Edges are fetched on first request and cached in the citations/works tables; citing lists are refreshed after 90 days, references never change. Cited works that are themselves in the index come back with their record_id, and search hits carry cited_by_count once known.

Verified 2026-08-27: the Oregon DMV DIP evaluation (dot:21848, no DOI, no date in ROSA-P) resolved by title and lists 6 citing works, among them Iowa's DIP evaluation and the NJ recidivism study; Lynn's 1982 Virginia 24-month report is cited by the 2003 Cochrane review of post-licence driver education (pubmed:12917984, in the index); a 2020 Seoul elderly-pedestrian paper has 57 references, 19 in the index. OpenAlex's coverage of NTL's 10.21949/… DOIs is thin: a bulk pass resolved 4,257 of the 15,493 ROSA-P records that carry one, and a random sample of the remainder found 1 in 50 known to OpenAlex. So most ROSA-P reports — DOI or not — depend on the title match, which accepts an exact normalised title, a prefix relation (edition or subtitle tails), or ≥ 0.8 token overlap, always with the year within ±1; when several OpenAlex entries fit (it holds duplicates of many reports), the most-cited one is taken as canonical. Grey literature that nobody has cited in indexed venues will still show zero; that is a property of the citation data, not of the index. OpenCitations and Semantic Scholar could be added as fallbacks in the same tables.

Semantic search (v0.4)

Keyword search is FTS5/BM25. Adding vectors turns search_reports into a hybrid search (BM25 and cosine fused by reciprocal rank) that finds records by meaning and across languages — an English query reaching Swedish, German, Spanish, Portuguese or Japanese records. Everything runs locally; no account, no GPU.

uv tool install "transport-lit[semantic]"   # adds fastembed (ONNX runtime), ~60 MB
transport-lit embed                         # default backend: fastembed, multilingual MiniLM-L12 (384-d, 220 MB model, one-time download)
transport-lit embed --backend ollama --model qwen3-embedding:8b     # opt-in: any Ollama embedding model, truncated to 1024-d
transport-lit search "programa de mejoramiento de conductores" --mode semantic

Semantic results are diversified by source: no single source may fill more than half of the requested results unless source/collection narrows the search (TRANSPORT_LIT_SEMANTIC_PER_SOURCE). Measured reason: CiNii is a third of the index and holds thousands of short English titles ("Pedestrian safety problems and countermeasures") that sit nearer a short query than any abstract-bearing record; the with/without-abstract cosine gap is only ~0.01, so this is corpus composition, not a length artifact, and a cap is the honest remedy. mode="semantic" is the specialist setting; hybrid stays the default.

Hybrid fusion weights the keyword list 1.0 and the semantic list 0.7 (TRANSPORT_LIT_SEMANTIC_WEIGHT), and a semantic-only candidate must clear cosine 0.5 (TRANSPORT_LIT_SEMANTIC_MIN); that keeps precise queries precise while mode="semantic" stays the recall / cross-language setting.

embed only processes records that have no vector yet, so after the first pass the weekly harvest adds seconds. Vectors live in $TRANSPORT_LIT_DATA_DIR/vectors/<backend-model>/ as a memory-mapped float16 matrix (342k × 384 ≈ 260 MB); search is a chunked dot product, no extension. The active vector set is recorded in the index, so search_reports(mode=…) uses whichever backend produced it: hybrid (default), keyword, or semantic; mode_used in every result says what ran, and it degrades to keyword when no vectors exist. harvest_status() reports backend, model, dimension and coverage.

Backends measured on 2026-08-26 on a 10-core Apple Silicon laptop, 256 real records (title + abstract): fastembed MiniLM-L12 ≈ 30 records/s on CPU (the CoreML provider is no faster); Ollama qwen3-embedding:0.6b ≈ 20/s (1024-d), qwen3-embedding:8b ≈ 1.4/s (4096-d, truncated to 1024). So the first full pass over 342k records is a one-time ~3 h with the default model; the weekly increment is seconds. Use --source to embed one source with a heavier model. Note the MiniLM model reads at most 128 tokens (title plus the first ~90 words of the abstract); Qwen reads the full 1,500-character window and gets the model's retrieval instruction prefix on queries. Most users should install a snapshot (below) and never run the full pass at all.

Snapshots: skip the harvest

uv tool install "transport-lit[semantic]"
transport-lit snapshot install https://github.com/aquistbe/transport-lit/releases/download/v0.4.0/transport-lit-2026-08.tar.gz
transport-lit mcp-config claude-desktop     # or install-claude-desktop --write

That is a complete, searchable install in minutes: 224k records (everything except CiNii and TRID) with vectors. transport-lit snapshot build <file.tar.gz> packs the SQLite index plus the active vectors; snapshot install <url-or-file> unpacks one into a fresh TRANSPORT_LIT_DATA_DIR, after which weekly incremental harvests keep it current (the snapshot carries the harvest bookkeeping, so harvest --source all knows where to resume). Snapshots leave out CiNii (its API terms require registration and are silent on redistribution) and TRID imports (TRB's terms); users harvest those themselves. Releases carry a snapshot when one was built.