Search your document archive

This guide is for a team with a mounted document folder that its operator may read and index. Archive Tools creates a local SQLite index of documents and images without changing the source files. You can search that index and investigate questions that need evidence from several files.

Choose a folder that contains work your team needs to search. Start with a small, approved part of the archive. Keep the database and run records outside the source folder. The plugin does not copy the source drive's access rules to the index, so protect the index under your organization's access and retention policy.

Install Archive Tools

On an Apple Silicon Mac, install mere.run and check model capacity. Archive Tools uses local text anonymization, image embedding, optical character recognition (OCR), and image captioning. The first caption can download its default checkpoint, so complete the first index while connected to the network.

mere.run model capabilities
mere.run model pull text-anonymize-privacy-filter
mere.run model pull vision-embed-qwen3-vl-2b
mere.run model pull vision-ocr-lighton
mere.run plugin install mere-archive-tools --yes
mere-archive-tools doctor

The doctor command checks command availability, document conversion, and SQLite full-text search. It does not confirm that every model is installed or index a file. For model storage, see Model management.

Index an approved folder

In the following commands, replace /Volumes/Shared/APPROVED_FOLDER with the mounted folder that your operator may index. Choose a work directory outside that folder. The umask setting limits access to newly created local files:

ARCHIVE_SOURCE="/Volumes/Shared/APPROVED_FOLDER"
ARCHIVE_WORK_DIR="$HOME/mere-run-archive"
umask 077
mkdir -p "$ARCHIVE_WORK_DIR"
chmod 700 "$ARCHIVE_WORK_DIR"
mere-archive-tools index \
  --source "$ARCHIVE_SOURCE" \
  --database "$ARCHIVE_WORK_DIR/archive.sqlite3" \
  --output-dir "$ARCHIVE_WORK_DIR/first-run" \
  --storage-tier safe-content

The safe-content tier stores short, reduced summaries, keywords, and embeddings—numeric representations used for similarity search. It does not store the complete extracted document body. It still retains file paths and derived information that may be sensitive. The plugin reduces personally identifiable information (PII) before embedding text, but that reduction cannot guarantee that every sensitive value is removed.

To check coverage and file errors, inspect the index statistics and run record:

mere-archive-tools stats --database "$ARCHIVE_WORK_DIR/archive.sqlite3"
cat "$ARCHIVE_WORK_DIR/first-run/run.json"

Resolve file errors before treating search results as complete. The plugin supports common text, Office, PDF, and image files. It rejects hosted OCR during document conversion, so a scanned PDF may appear as a file error. For source formats and storage tiers, see the Archive Tools plugin guide.

Search and check the source

Set ARCHIVE_QUESTION to a question your team needs answered from this folder. Replace the uppercase placeholder before running the command:

ARCHIVE_QUESTION="REPLACE_WITH_A_QUESTION_ABOUT_YOUR_ARCHIVE"
mere-archive-tools search \
  --database "$ARCHIVE_WORK_DIR/archive.sqlite3" \
  --query "$ARCHIVE_QUESTION"

Search results identify source paths and whether each file is available. Open the retrieved files and check that they cover the question. If a result points to an unavailable path, mount the source drive again. The index does not contain a copy of the original file.

If the question requires several searches, install Pi, the local agent runner, and the default investigation model. Follow the support check from mere.run model capabilities; do not bypass it to fit a model on an undersized Mac:

mere.run agent onboard --install-pi
mere.run model pull text-chat-bonsai-27b-2bit
mere-archive-tools investigate \
  --database "$ARCHIVE_WORK_DIR/archive.sqlite3" \
  --question "$ARCHIVE_QUESTION" \
  --diagnostics "$ARCHIVE_WORK_DIR/investigation-metrics.json"

The investigator can make up to four searches with five results per search by default. It returns source-linked claims and unresolved points. Check each claim in the cited file before using the answer. A citation confirms that a path was returned by search; it does not prove that the file supports the claim. If the investigation model does not fit your Mac, use search without the investigator.

Maintain and evaluate the index

For an interrupted index run, inspect the manifest before resuming it:

mere-archive-tools resume "$ARCHIVE_WORK_DIR/first-run/run.json"

To refresh the archive after source files change, run index again with the same database and a new output directory. The plugin reuses identical content and removes index records for paths that disappeared. It does not delete source files.

Check retrieval against questions whose answers and source files your team already knows. Record missing files, incorrect matches, claim errors, and elapsed time. Review the index's access, backup, and retention settings before you let more people query it. Image mode embeds reduced captions and OCR by default; --image-index visual also embeds source pixels and needs a separate privacy review.