nasindex Documentation

One indexing cycle

The wizard shows these steps by their German names.

The first cycle over a large archive takes hours. After that, a cycle over an unchanged archive takes only seconds to minutes.

What the path reveals

In any archive the first folder level counts as the project and the second as the phase. The year comes from the path or file name. If a document in a construction project archive sits under a folder «Aufträge» (orders), nasindex also recognises BKP code, trade, award stage and document class. From 24031-MST / 1 Aufträge / 22810_Sonnenschutz Südfassade / 2 Vergabe / 1 Angebote / Offerte_250312.pdf it reads:

BKP is the Swiss construction cost classification. Duplicates are merged by content. The copy that wins has a document class, then a trade, then lies outside backup, copy or dispatch folders, then has the shorter path. The year comes from the path or file name, otherwise from the file date. It is good enough for date-range filters, not for statements like «awarded in year X».

Requirements

• Docker with Docker Compose v2. Check with docker compose version.
• At least 5 GB free in the installation folder, plus space for the index, about a tenth of the archive size.
• Tailscale on the computer running nasindex and on every workstation using Claude.
• On the workstations Claude Desktop with Node.js, or Claude Code.
• In cifs mode: a NAS user with read-only access to the archive share. nasindex never needs an administrator account.

Starting

Unpack the package, do not rename or move the nasindex folder, then inside it:

sh installieren.sh

On the Mac a double-click on Installieren.command works too. If Docker needs administrator rights, as is usual on a Synology NAS, the script asks for the password itself.

Seed: catalogue from another computer

A more powerful computer builds the index once in full. Its catalogue consists of four entries: catalog.jsonl.gz, catalog.json, text/ and meta/. Inside the container they are under /daten:

mkdir nasindex-katalog
docker cp nasindex:/daten/catalog.jsonl.gz nasindex-katalog/
docker cp nasindex:/daten/catalog.json nasindex-katalog/
docker cp nasindex:/daten/text nasindex-katalog/
docker cp nasindex:/daten/meta nasindex-katalog/

Copy the folder to the target system and give it to the wizard when it asks for the seed path. On the first start nasindex loads it, visible in the log as a line seed: …. After that the reconciliation usually reports «0 neu, 0 geändert» (0 new, 0 changed).

What it creates

All files have permissions 600, only the owner can read them. The wizard overwrites existing files only after asking. In cifs mode it checks the credentials against the NAS before starting, and asks again up to three times on failure.

Progress and running it again

After the start the wizard shows the first cycle live, for example Schritt 5 von 7: Texte auslesen - 1234 von 20000 (6 %), noch ~3 h 10 min. Ctrl-C only ends the display, indexing continues. To see the status again later:

sh installieren.sh

The script then asks «Die Fragen neu beantworten?» (answer the questions again?). Enter means no: it only checks and shows the status.

Questions this answers

meta_only means no full text, for example for CAD drawings and images. They can still be found by path and file name. missing means the file has disappeared from the archive.

Updating

An update is a new package. Unpack it in the same place and let it overwrite existing files. .env and the credentials are not in the package and stay. Then run sh installieren.sh and answer «Die Fragen neu beantworten?» with Enter (no). The index is kept.

Providers

Digital Apes has only tested anthropic with the default model. With openai and eigener, the model must follow a JSON schema and fit documents of up to 500,000 characters, about 125,000 tokens.

Procedure

docker compose exec nasindex python3 /app/nasindex.py felder --trockenlauf
docker compose exec nasindex python3 /app/nasindex.py felder --limit 15
docker compose exec nasindex python3 /app/nasindex.py felder

The dry run (--trockenlauf) shows count, characters and, with anthropic, the estimated cost, without sending a document. Reference value: 541 contracts with 14.7 million characters cost around 13 to 15 US dollars. --limit 15 gives a small set for proofreading. An interrupted run continues on the next call. Do not start it in parallel with a running indexing cycle.

Pre-check with your own model

Before every run nasindex asks the server for the context size and sends a probe with a made-up mini contract. That shows in advance whether the model keeps to the format. Documents that are too large are left for a model with more context. If the server does not report a context size (Ollama, LM Studio, llama.cpp), set it in the wizard or as EXTRAKTION_KONTEXT. Ollama silently truncates inputs that are too long.

Field rules

• Contract number only if the document calls it a contract, order, purchase, addendum or additional-cost number. Never project, object, cost-centre or BKP numbers.
• The contractor is never the architect, specialist planner, site management or engineering firm, except in that planner's own contract. Nor a delivery or invoice address.
• Contractor's signatory only from the contractor's company.
• Currency only with a stated amount, never from an IBAN.
• Phone and email only if they clearly belong to the contractor.
• Optionally a client name: that name is then never taken as the contractor.

A rule change only affects contracts read afterwards. felder --force reads all of them again and costs as much as a full run.

Resources: The Compose files set 4 CPUs and 4 GB of memory. Rule of thumb: 1 GB per file read at the same time plus 1 GB base load. An invalid value stops the start with a message that names only the setting.

Folder removed on purpose

The safety guard prevents a briefly missing network drive from emptying the index. If the drop is intended, confirm it once:

docker exec nasindex python3 /app/nasindex.py scan --verschwunden-ok --root /archiv --index /daten --cache /daten --rolle indexer

Messages from archiv_status

Exit codes of build

If you have questions, send Digital Apes the output of docker compose logs --tail 200 nasindex. It contains no credentials.