Question 1

Why does Paperless-ngx fail with 'input (681 tokens) is too large to process. increase the physical batch size' when using llama.cpp for embeddings?

Accepted Answer

The error occurs because the embedding backend (llama.cpp) receives an input larger than its configured physical batch size (`-b 512`). To fix, increase the batch size in the llama.cpp command to at least 1024 (e.g., `-b 1024`). Alternatively, use an embedding model that supports larger input sizes. Paperless-ngx currently hardcodes `RAG_CHUNK_SIZE` to 1024, which can cause this issue for models with smaller context windows; a future update may allow adjusting this value.

Question 2

How to fix paperless.W001 when PAPERLESS_DB_OPTIONS JSON is ignored and the database timeout is not applied?

Accepted Answer

In paperless-ngx v3.3.0, do not use the JSON form shown in the warning:

```bash
# Not parsed in v3.3.0; the option is silently ignored
PAPERLESS_DB_OPTIONS='{"timeout": "60"}'
```

Use `key=value` syntax instead. For SQLite:

```bash
PAPERLESS_DB_OPTIONS=timeout=60
```

For PostgreSQL, the connection-timeout option is `connect_timeout`, not `timeout`:

```bash
PAPERLESS_DB_OPTIONS=connect_timeout=60
```

With Docker:

```bash
docker run -e PAPERLESS_DB_OPTIONS=timeout=60 ...
# PostgreSQL:
docker run -e PAPERLESS_DB_OPTIONS=connect_timeout=60 ...
```

Remove `PAPERLESS_DB_TIMEOUT` completely after migrating. Do not set it to an empty value (`PAPERLESS_DB_TIMEOUT=`), because v3.3.0 fails with `ValueError: invalid literal for int() with base 10: ''`. The old variable still works in v3.3.0 despite the warning saying it would be removed in v3.2, but using the correct `PAPERLESS_DB_OPTIONS` form clears the warning and applies the setting.

Question 3

Why are tags not saving when using bulk edit in Paperless-ngx?

Accepted Answer

Known issue when running Paperless-ngx under ASGI (e.g., Granian) with PostgreSQL and psycopg3. The bulk modify_tags operation silently fails because the M2M write occurs in a different database transaction that never commits. Workaround: Edit tags from the document detail view (PATCH /api/documents/{id}/) where the operation succeeds. Alternatively, if possible, switch to a WSGI server (e.g., Gunicorn) to avoid the async/sync boundary problem. Track the GitHub issue #12699 for a permanent fix.

Question 4

How do I fix Paperless-ngx AI suggestions returning HTTP 500 with thinking-capable Ollama models such as gpt-oss:20b or qwen3:14b?

Accepted Answer

The failure occurs because the Ollama/LlamaIndex integration handles the per-call `think=False` option incorrectly, leaving `message.content` empty; Paperless-ngx then fails while parsing the expected JSON response.

Temporary workaround in Paperless-ngx 3.1.3:
1. Open the AI/LLM settings.
2. Change the Ollama model from `gpt-oss:20b` or `qwen3:14b` to a non-thinking model such as `llama3.1:8b` or `gemma3:12b`.
3. Save and retry Suggest.

When calling Ollama directly, omit the `think` parameter instead of sending `"think": false`:

```bash
curl http://<ollama>:11434/api/chat -d '{
  "model": "gpt-oss:20b",
  "messages": [{"role":"user","content":"Return JSON with a field name set to test."}],
  "stream": false,
  "format": {"type":"object","properties":{"name":{"type":"string"}},"required":["name"]}
}'
```

Permanent fix: upgrade to a Paperless-ngx release containing the updated Ollama/LlamaIndex fix `fix(ollama): honor per-call think=False`. For Docker:

```bash
docker compose pull
docker compose up -d
```

After upgrading, switch back to `gpt-oss:20b` or `qwen3:14b` and test AI suggestions again.

Question 5

How to fix "Error listing search results" on saved views with advanced queries after upgrading Paperless-ngx to 3.0?

Accepted Answer

Paperless-ngx 3.0 switched its search backend (Whoosh → Tantivy), which breaks some previously valid advanced query syntax in saved views. Two workarounds:

1. Replace the `AND NOT field:value` construction with the negation operator `-field:value`. For example, change `AND NOT title:2024` to `AND -title:2024`.

2. Remove wildcard/regex character-class patterns like `title:*202[0-3]` — wildcards are not yet supported in the new search backend. Enumerate the values explicitly instead.

Example rewritten query that parses and filters correctly:
```
tag:steuer AND tag:valentin AND (title:2025 OR (-title:2024 AND -title:2023 AND -title:2022 AND -title:2021 AND -title:2019 AND -title:2018 AND created:2025))
```

Edit each affected Saved View and update its query string accordingly. Full support for the old `AND NOT` behavior and wildcards depends on an upstream Tantivy fix and future Paperless-ngx releases.

Paperless-ngx

Core Features

What It Can't Do

Use Cases

Tags

Getting Started

Similar Projects