Mirroring Bioconductor and building a DuckDB code index for AI models

The setup described here ran on a virtual machine on Jetstream2, referred to below as “this instance”. Nothing in it depends on Jetstream2: any cloud instance or local machine with enough disk space will do.

This instance does two things. It keeps a local copy of Bioconductor’s source packages for release 3.23 and devel 3.24. It also runs scip-r, written by Sean Davis, over every package, which produces an index of symbols, references and call relationships, and packages that index as a DuckDB database that an AI model can query.

This post walks through how the mirror and index were set up, and how bioc-index.sh turns the index into a database.¹ The mirror steps follow the mirror how-to. The scripts used here can be downloaded from the gist.

Where things stand

As of 12 November 2026:

• Source mirror: packages/3.23/bioc/src/contrib (8.2 GB) and packages/3.24/bioc/src/contrib (6.9 GB), under ~/bioconductor_repositories.
• Index: out/ in each version directory, with a parquet/ tree and a bioc323.duckdb or bioc324.duckdb view database.
• Bundle: bioc324-index-20261104.tar.gz (118.6 MB) with a .sha256 file, built on 4 November 2026 from 3.24.

Step 1: read the manifest

Each version and repository has a manifest listing every object, with its size and MD5. Filter the bioc repository’s manifest down to the source tarballs and index files:

cd ~/bioconductor_repositories
curl -fsS https://bioconductor.org/api/v1/manifest/3.24/bioc.tsv.gz \
  | gunzip | cut -f1 | grep '/src/contrib/' > files-3-24.txt

The same command with 3.23 gives files-3-23.txt. The 3.23 list has 2,763 paths and the 3.24 list has 2,456.

Step 2: copy with rclone

Copy the listed files with rclone:

rclone copy --http-url https://bioconductor.org :http: \
  ~/bioconductor_repositories --files-from files-3-24.txt --no-traverse

The how-to’s mirror-sync.sh runs this copy and then recreates the symlinks. This instance ran the copy by hand.

Paths in --files-from are relative to the destination, so each file lands under packages/<version>/bioc/src/contrib/. Every path in both lists is present on disk.

The list counts are higher than the tarball counts (2,763 against 2,758, and 2,456 against 2,451) because each list also includes non-tarball files. For 3.24 those are PACKAGES, PACKAGES.gz, PACKAGES.rds and two Meta/*.rds files.

What this copy leaves out

• Scope is source only. The mirror has the bioc repository’s src/contrib and nothing else: no bin/, no data-annotation, data-experiment, workflows or books. The how-to leaves that choice to you, and this instance picked source only.
• No release or devel symlinks. packages/ contains only 3.23 and 3.24. The how-to says URLs that go through those links return 404 without them. mirror-sync.sh recreates them; this copy did not.
• The copy is a snapshot. It dates from 4 November 2026. The manifest is regenerated hourly, and this user has no cron job, so nothing refreshes it. The how-to suggests a daily entry: 17 3 * * * DEST=/srv/bioc /usr/local/bin/mirror-sync.sh.

Step 3: build the index with scip-r

Install scip-r as a uv tool, as in its README:

uv tool install "scip-r[export] @ git+https://github.com/seandavi/scip-r"

For those interested, I submitted a feature request to scip-r as pull request #52 to add full function signatures in the index.

Run the batch in each version’s bioc directory:

cd ~/bioconductor_repositories/packages/3.24/bioc
scip-r batch src/contrib -o out --jobs 14 --export parquet --hive

The --resolve flag records the installed R package versions. It needs R on the machine, so add it only if R is installed (sudo apt install r-base-core on Debian or Ubuntu). This index was built without it: none of the rows in out/summary.jsonl has a resolve_run_id, so it is a static index.

The output is one directory per package under out/, a summary.jsonl with one line per package, and out/parquet/, with one directory per table. --hive partitions each table by index_package and index_version.

For 3.24, summary.jsonl has 2,451 rows, all with status ok. The index has symbols for 2,431 packages. Six of the packages with no symbols are AHMassBank, BiocVersion, alabaster.schemas, alabaster, assorthead and convert. Why the other 14 have no symbols is not yet explained.

Step 4: DuckDB views over the Parquet

Parquet is the store and DuckDB reads it through views. ~/make_duckdb.sh creates them:

~/make_duckdb.sh out out/bioc324.duckdb

Each view is a read_parquet over one table directory, for example:

CREATE OR REPLACE VIEW symbols AS
  SELECT * FROM read_parquet('/abs/path/out/parquet/symbols/**/*.parquet',
                             hive_partitioning=true);

The views store absolute paths, so the database copy in the bundle only works on the machine that built it. Install rebuilds the views against the local Parquet files.

The schema is described in scip-r’s docs/parquet-schema.md. The six tables are:

For 3.24 the database holds 12,307,301 occurrences, 338,790 external symbols, 36,721 documents and 27,175 relationships.

Step 5: bundle on the VM

bioc-index.sh is in the working directory. Bundle 3.24 from there:

cd ~/bioconductor_repositories
./bioc-index.sh bundle 3.24

The script takes 3.24 and works out the names: the tag is bioc324 and the bundle is bioc324-index-<date>.tar.gz. Its default source is packages/3.24/bioc/out, and it writes into the current directory. The bundle contains parquet/ and bioc324.duckdb. The database is included for reference only, because install rebuilds it. The script also writes a .sha256 file next to the tarball.

Step 6: install where the model runs

Copy the tarball and its .sha256 to the machine that will serve the index. There, run:

./bioc-index.sh install bioc324-index-20261104.tar.gz

The install verifies the checksum, checks that the duckdb CLI is on the path, extracts the bundle into ~/bioc-index/<bundle name> unless you pass a destination, and recreates the views against the local Parquet files. It then prints package and symbol counts. The output below is from the Jetstream2 VM, where the default user is exouser. On your own machine, ~/bioc-index is under your home directory instead:

bioc324-index-20261104.tar.gz: OK
Installed to /home/exouser/bioc-index/bioc324-index-20261104
│     2431 │  191331 │
Database: /home/exouser/bioc-index/bioc324-index-20261104/bioc324.duckdb

That is 305 MB on disk, and it took about six seconds.

If you use the snap-packaged duckdb, pass a destination under $HOME, because the snap cannot see /tmp.

Step 7: query it

The queries in this step show what the index can answer, and they are useful for checking that the install worked. In day-to-day use you would not write them yourself. The AI model writes and runs them, as described in the next section. Each query below is paired with a prompt you might give the model to get it.

Open the database read-only:

duckdb -readonly ~/bioc-index/bioc324-index-20261104/bioc324.duckdb

The packages with the most function symbols:

Which five Bioconductor packages define the most functions?

SELECT index_package, count(*) AS n_symbols
FROM symbols WHERE is_function
GROUP BY 1 ORDER BY 2 DESC LIMIT 5;

This returns ChemmineOB (4,566), mia (1,392), igblastr (1,353), xcms (1,235) and S4Vectors (1,049).

A call graph for limma, using the GROUP BY caller, symbol pattern from the schema doc:

In limma, which functions call which other functions most often?

SELECT caller, symbol, count(*) AS n
FROM occurrences
WHERE caller IS NOT NULL AND NOT is_definition AND index_package = 'limma'
GROUP BY ALL ORDER BY n DESC LIMIT 4;

The top rows show printtipWeights() calling base::sum() 136 times. The caller and symbol columns hold raw SCIP strings such as scip-r . base . sum()., and local N marks a file-local symbol. For a clean function-level graph, add AND NOT is_local and use the parsed package and name columns instead of splitting the strings.

Counting how often each function from another package is called:

Across Bioconductor, which functions from other packages are called most often, and by how many packages?

SELECT package, name, count(*) AS calls,
       count(DISTINCT index_package) AS n_pkgs
FROM occurrences
WHERE NOT is_local AND NOT is_definition AND is_function
  AND package <> index_package
GROUP BY ALL ORDER BY calls DESC LIMIT 10;

Including base R, base::c leads with 238,725 calls from 2,376 packages, followed by base::length and base::return. Excluding base R, which you can do by adding AND package NOT IN ('base', 'methods', 'stats', 'utils', 'grid', 'graphics', 'grDevices', 'tools', 'parallel') (and the other base packages) to the WHERE clause, ggplot2::aes leads with 5,794 calls from 465 packages, followed by dplyr::mutate and dplyr::filter. The prompt for that version:

Same question, but leave out base R packages.

Two things to know about these counts:

• return is a function call in R, so it ranks high. Drop it if you only want library functions.
• Ranking by n_pkgs instead of calls finds functions that many packages share, rather than functions that one package calls heavily.

Giving it to a model

A reasonable setup gives the model two things: the scip-r schema doc, and one tool that runs read-only SQL against the database. With the schema doc in context, the model can write its own queries. That tool isn’t built yet, and it is the next step.

Things to change next time

1. Use mirror-sync.sh, or make the release and devel symlinks by hand, so URLs through those names work.
2. Schedule the sync. The copy dates from 4 November 2026, and the manifest is regenerated every hour.
3. Install under $HOME when you use the snap duckdb.
4. Decide whether the index covers source only, as it does here, or the binaries too.
5. bioc-index.sh hard-codes SRC_ROOT to /home/exouser/bioconductor_repositories/packages, the path on the Jetstream2 VM. On another machine, or as a user other than exouser, either edit SRC_ROOT or pass the index directory as the second argument: ./bioc-index.sh bundle 3.24 ~/bioconductor_repositories/packages/3.24/bioc/out.