paperless-gpt#

paperless-gpt is a sidecar container that adds LLM-powered auto-tagging, titling, correspondent assignment, and date extraction to Paperless-NGX. It runs alongside the Paperless stack on ereshkigal and delegates inference to Ollama on mokou.

Architecture#

```mermaid flowchart LR subgraph paperless_default[“paperless_default Podman network”] web[“paperless-web\n:8000 (internal)”] gpt[“paperless-gpt\n:8080 (internal)\n:8013 (host)”] end nginx[“nginx\n:8014 HTTPS”] mokou[“Ollama on mokou\n:11434”]

gpt -->|"REST API + token"| web
gpt -->|"LLM / VLM calls"| mokou
nginx -->|"proxy"| gpt

```

paperless-gpt and paperless-web share the paperless_default Podman network, so they communicate by container hostname (webserver) rather than via the host.

How it works#

Processing is triggered by tags on documents:

The recommended workflow: create a Paperless workflow (Admin → Workflows) with trigger “Document Added” and action “Assign tag paperless-gpt-auto”. All newly ingested documents are then processed automatically in the background without hitting the 60s UI timeout.

For already-existing documents, tag them manually with paperless-gpt and use the UI to review and apply suggestions.

Module options#

Configured under tsunaminoai.docPipeline.paperlessGpt:

The vision model for OCR is inherited from docPipeline.ollamaModel (qwen2.5-vl:7b).

SOPS secret#

The Paperless API token is stored in SOPS and injected into the container via a systemd EnvironmentFile. A sops.templates entry formats it as KEY=VALUE:

sops.secrets."paperless/api-token" = {};

sops.templates."paperless-gpt-env" = {
  content = "PAPERLESS_API_TOKEN=${config.sops.placeholder."paperless/api-token"}";
  mode = "0400";
};

The podman container gets --env=PAPERLESS_API_TOKEN (no value — this form reads from the process environment, which systemd populates from the EnvironmentFile).

To create the secret:

# On the machine with your SOPS key
sops secrets/ereshkigal.yaml
# Add:  paperless/api-token: <token from Paperless Admin → Tokens>

Granian / IPv6 binding#

Paperless-NGX uses the Granian ASGI server, which binds IPv6-only (:::8000) by default. paperless-gpt resolves webserver to an IPv4 address inside the Podman network, causing connection refused errors.

Fix: set GRANIAN_HOST=0.0.0.0 in the paperless-web container environment (not just PAPERLESS_BIND_ADDR — Granian reads its own env var directly).

Tag creation#

CREATE_NEW_TAGS=true allows paperless-gpt to create tags in Paperless that don’t yet exist. Without this, it can only assign pre-existing tags.

To bootstrap a useful starting set, use the Paperless API:

set TOKEN (ssh ereshkigal sudo cat /run/secrets/paperless/api-token)
set BASE "https://ereshkigal.armadillo-banfish.ts.net:8012"

set tags tax-return tax-document bank-statement pay-stub invoice receipt \
  investment loan contract agreement legal identity medical prescription \
  employment insurance vehicle education government correspondence personal

for tag in $tags
  curl -sf -o /dev/null -w "%{http_code}  $tag\n" \
    -X POST "$BASE/api/tags/" \
    -H "Authorization: Token $TOKEN" \
    -H "Content-Type: application/json" \
    -d "{\"name\": \"$tag\"}"
end

Tag prompt#

The tag prompt is editable in the paperless-gpt UI (Settings → Prompts). With CREATE_NEW_TAGS=true, update it to allow new tag suggestions:

I will provide you with the content and the title of a document.
Your task is to select appropriate tags for the document.
Prefer tags from the provided list when they fit. You may also suggest new
tags not in the list if none of the existing tags adequately describe the document.
Respond only with the selected tags as a comma-separated list.
New tags should be short, lowercase, and general enough to apply to multiple documents.

<available_tags>
{{.AvailableTags | join ", "}}
</available_tags>

<title>{{.Title}}</title>
<content>{{.Content}}</content>

Performance notes#

• Processing time: ~30s–6min per document depending on page count and whether vision OCR runs. Background auto-processing has no timeout; the UI endpoint times out at 60s.
• PDF_SKIP_EXISTING_OCR=true: Paperless+tesseract already runs OCR on ingest, so paperless-gpt skips re-running it on documents that already have a text layer.
• TOKEN_LIMIT=1000: Truncates document content to ~750 words before sending to the LLM. Sufficient for metadata extraction; prevents context overflow.

Troubleshooting#

connection refused on startup — paperless-gpt started before paperless-web was ready. Normal; it retries every 20s and recovers automatically.

Token not reaching container — verify the sops template path matches the EnvironmentFile path in the systemd service, and that --env=PAPERLESS_API_TOKEN (without a value) is in extraOptions.

Processing very slow / timing out — check that Ollama is using the GPU: nvidia-smi on mokou should show nonzero GPU utilisation during inference. If not, see Ollama — The CUDA SM 6.1 problem.