pleroma/docs/configuration/search.md

# Configuring search

{! backend/administration/CLI_tasks/general_cli_task_info.include !}

## Built-in search

To use built-in search that has no external dependencies, set the search module to `Pleroma.Activity`:

> config :pleroma, Pleroma.Search, module: Pleroma.Search.DatabaseSearch

While it has no external dependencies, it has problems with performance and relevancy.

## QdrantSearch

This uses the vector search engine [Qdrant](https://qdrant.tech) to search the posts in a vector space. This needs a way to generate embeddings and uses the [OpenAI API](https://platform.openai.com/docs/guides/embeddings/what-are-embeddings). This is implemented by several project besides OpenAI itself, including the python-based fastembed-server found in `supplemental/search/fastembed-api`.

The default settings will support a setup where both the fastembed server and Qdrant run on the same system as pleroma. To use it, set the search provider and run the fastembed server, see the README in `supplemental/search/fastembed-api`:

> config :pleroma, Pleroma.Search, module: Pleroma.Search.QdrantSearch

Then, start the Qdrant server, see [here](https://qdrant.tech/documentation/quick-start/) for instructions.

You will also need to create the Qdrant index once by running `mix pleroma.search.indexer create_index`. Running `mix pleroma.search.indexer index` will retroactively index the last 100_000 activities.

### Indexing and model options

To see the available configuration options, check out the QdrantSearch section in `config/config.exs`.

The default indexing option work for the default model (`snowflake-arctic-embed-xs`). To optimize for a low memory footprint, adjust the index configuration as described in the [Qdrant docs](https://qdrant.tech/documentation/guides/optimize/). See also [this blog post](https://qdrant.tech/articles/memory-consumption/) that goes into detail.

Different embedding models will need different vector size settings. You can see a list of the models supported by the fastembed server [here](https://qdrant.github.io/fastembed/examples/Supported_Models), including their vector dimensions. These vector dimensions need to be set in the `qdrant_index_configuration`. 

E.g, If you want to use `sentence-transformers/all-MiniLM-L6-v2` as a model, you will not need to adjust things, because it and `snowflake-arctic-embed-xs` are both 384 dimensional models. If you want to use `snowflake/snowflake-arctic-embed-l`, you will need to adjust the `size` parameter in the `qdrant_index_configuration` to 1024, as it has a dimension of 1024.

When using a different model, you will need do drop the index and recreate it (`mix pleroma.search.indexer drop_index` and `mix pleroma.search.indexer create_index`), as the different embeddings are not compatible with each other.

## Meilisearch

Note that it's quite a bit more memory hungry than PostgreSQL (around 4-5G for ~1.2 million
posts while idle and up to 7G while indexing initially). The disk usage for this additional index is also
around 4 gigabytes. Like [RUM](./cheatsheet.md#rum-indexing-for-full-text-search) indexes, it offers considerably
higher performance and ordering by timestamp in a reasonable amount of time.
Additionally, the search results seem to be more accurate.

Due to high memory usage, it may be best to set it up on a different machine, if running pleroma on a low-resource
computer, and use private key authentication to secure the remote search instance.

To use [meilisearch](https://www.meilisearch.com/), set the search module to `Pleroma.Search.Meilisearch`:

> config :pleroma, Pleroma.Search, module: Pleroma.Search.Meilisearch

You then need to set the address of the meilisearch instance, and optionally the private key for authentication. You might
also want to change the `initial_indexing_chunk_size` to be smaller if you're server is not very powerful, but not higher than `100_000`,
because meilisearch will refuse to process it if it's too big. However, in general you want this to be as big as possible, because meilisearch
indexes faster when it can process many posts in a single batch.

> config :pleroma, Pleroma.Search.Meilisearch,
>    url: "http://127.0.0.1:7700/",
>    private_key: "private key",
>    initial_indexing_chunk_size: 100_000

Information about setting up meilisearch can be found in the
[official documentation](https://docs.meilisearch.com/learn/getting_started/installation.html).
You probably want to start it with `MEILI_NO_ANALYTICS=true` environment variable to disable analytics.
At least version 0.25.0 is required, but you are strongly advised to use at least 0.26.0, as it introduces
the `--enable-auto-batching` option which drastically improves performance. Without this option, the search
is hardly usable on a somewhat big instance.

### Private key authentication (optional)

To set the private key, use the `MEILI_MASTER_KEY` environment variable when starting. After setting the _master key_,
you have to get the _private key_, which is actually used for authentication.

=== "OTP"
    ```sh
    ./bin/pleroma_ctl search.meilisearch show-keys <your master key here>
    ```

=== "From Source"
    ```sh
    mix pleroma.search.meilisearch show-keys <your master key here>
    ```

You will see a "Default Admin API Key", this is the key you actually put into your configuration file.

### Initial indexing

After setting up the configuration, you'll want to index all of your already existing posts. Only public/unlisted posts are indexed.  You'll only
have to do it one time, but it might take a while, depending on the amount of posts your instance has seen. This is also a fairly RAM
consuming process for `meilisearch`, and it will take a lot of RAM when running if you have a lot of posts (seems to be around 5G for ~1.2
million posts while idle and up to 7G while indexing initially, but your experience may be different).

The sequence of actions is as follows:

1. First, change the configuration to use `Pleroma.Search.Meilisearch` as the search backend
2. Restart your instance, at this point it can be used while the search indexing is running, though search won't return anything
3. Start the initial indexing process (as described below with `index`),
   and wait until the task says it sent everything from the database to index
4. Wait until everything is actually indexed (by checking with `stats` as described below),
   at this point you don't have to do anything, just wait a while.

To start the initial indexing, run the `index` command:

=== "OTP"
    ```sh
    ./bin/pleroma_ctl search.meilisearch index
    ```

=== "From Source"
    ```sh
    mix pleroma.search.meilisearch index
    ```

This will show you the total amount of posts to index, and then show you the amount of posts indexed currently, until the numbers eventually
become the same. The posts are indexed in big batches and meilisearch will take some time to actually index them, even after you have
inserted all the posts into it. Depending on the amount of posts, this may be as long as several hours. To get information about the status
of indexing and how many posts have actually been indexed, use the `stats` command:

=== "OTP"
    ```sh
    ./bin/pleroma_ctl search.meilisearch stats
    ```

=== "From Source"
    ```sh
    mix pleroma.search.meilisearch stats
    ```

### Clearing the index

In case you need to clear the index (for example, to re-index from scratch, if that needs to happen for some reason), you can
use the `clear` command:

=== "OTP"
    ```sh
    ./bin/pleroma_ctl search.meilisearch clear
    ```

=== "From Source"
    ```sh
    mix pleroma.search.meilisearch clear
    ```

This will clear **all** the posts from the search index. Note, that deleted posts are also removed from index by the instance itself, so
there is no need to actually clear the whole index, unless you want **all** of it gone. That said, the index does not hold any information
that cannot be re-created from the database, it should also generally be a lot smaller than the size of your database. Still, the size
depends on the amount of text in posts.

## ParadeDB

[ParadeDB](https://www.paradedb.com/) is a Postgres extension that provides BM25 full-text search. In Pleroma, it can be used as an
external search backend by pointing Pleroma at a separate Postgres instance with the `pg_search` extension installed.
Pleroma's integration suite tests `pg_search` 0.24.3 on PostgreSQL 15.

Pleroma will maintain a small `pleroma_search_documents` table in that database (via the existing search indexing queue) and run search
queries against it.
The indexed text includes status content, content warnings, and attachment descriptions.

### Configuration

Set the search module to `Pleroma.Search.ParadeDB` and configure the ParadeDB database URL. These settings must be in static/runtime config
(or `PARADEDB_DATABASE_URL`) so the dedicated ParadeDB Repo is started with the application. They cannot be migrated to or changed through
ConfigDB.

> config :pleroma, Pleroma.Search, module: Pleroma.Search.ParadeDB
>
> config :pleroma, Pleroma.Search.ParadeDB,
>   url: System.get_env("PARADEDB_DATABASE_URL"),
>   table: "pleroma_search_documents",
>   fuzzy_distance: 0
>
> config :pleroma, Pleroma.Search.ParadeDB.Repo,
>   pool_size: 2,
>   prepare: :unnamed

When connecting over an untrusted network, enable certificate verification in the Repo configuration. For example:

> config :pleroma, Pleroma.Search.ParadeDB.Repo,
>   ssl: [
>     verify: :verify_peer,
>     cacertfile: "/etc/pleroma/paradedb-ca.pem",
>     server_name_indication: ~c"paradedb.example.com"
>   ]

### Initial rollout

The `pg_search` extension must already be available in the database. Creating the extension normally requires an administrative database
role. The Pleroma runtime role requires `CONNECT` on the database, `USAGE` on the table's schema and the `pdb` schema, and `SELECT`, `INSERT`, `UPDATE`,
and `DELETE` on the generated table. For example, after creating the table as its owner:

```sql
GRANT CONNECT ON DATABASE paradedb TO pleroma_search;
GRANT USAGE ON SCHEMA public, pdb TO pleroma_search;
GRANT SELECT, INSERT, UPDATE, DELETE ON TABLE pleroma_search_documents TO pleroma_search;
```

Use this rollout order to prevent jobs from being handled by the previous backend:

1. Install `pg_search` 0.24.3 on PostgreSQL 15 or another deliberately tested combination.
2. Stop Pleroma so no search indexing jobs are running.
3. Put the ParadeDB backend in static configuration.
4. Create the table and BM25 index in a one-shot offline process using `PARADEDB_DATABASE_URL` with the table owner/setup credential.
5. Apply the runtime grants shown above, set the service's `PARADEDB_DATABASE_URL` to the restricted runtime credential, and start Pleroma.
6. Start the initial backfill and monitor the `search_indexing` queue.

Create the table and BM25 index once with:

=== "OTP"
    ```sh
    PARADEDB_DATABASE_URL='OWNER_URL' PLEROMA_CTL_RPC_DISABLED=true \
      ./bin/pleroma_ctl search.indexer create_index
    ```

=== "From Source"
    ```sh
    PARADEDB_DATABASE_URL='OWNER_URL' mix pleroma.search.indexer create_index
    ```

The task currently prints an error without returning a failing process status. Verify that it prints `Index created` before continuing.

### Backfill

Run an initial indexing pass to enqueue existing posts:

=== "OTP"
    ```sh
    ./bin/pleroma_ctl search.indexer index
    ```

=== "From Source"
    ```sh
    PARADEDB_DATABASE_URL='RUNTIME_URL' mix pleroma.search.indexer index
    ```

The command enqueues up to 100,000 Create activities by default. It uses keyset pagination, so new activities arriving during the backfill
do not shift the remaining pages. The defaults and accepted options are:

* `--limit`: maximum activities to enqueue in this run; defaults to `100000` and must be positive
* `--step`: activities fetched from Pleroma's database per page; defaults to `1000` and must be positive
* `--chunk`: jobs inserted into Oban per batch; defaults to `100` and must be positive
* `--before`: resume before the checkpoint activity ID printed by an earlier run; the cursor is exclusive

The checkpoint is the last activity enqueued, not confirmation that its job succeeded. Repeat the command with its last `--before`
checkpoint only when it reports that more activities remain. Inspect failed or discarded Oban jobs as well as waiting jobs before declaring
the backfill complete.

The same operation can be run from an attached console while Pleroma is online:

```elixir
Pleroma.Search.Backfill.run(limit: 10_000)
```

The returned `next_cursor` can be supplied as `before` when `exhausted` is `false`. The attached API also accepts `limit: :infinity`, but this
can create an unbounded number of Oban rows; bounded runs are safer on large instances. An `on_page` callback may be supplied for progress
reporting.

Note: Like the other external search backends, only public/unlisted Notes are indexed.

Set `fuzzy_distance` to `1` or `2` to allow typo-tolerant matching (`2` is the maximum ParadeDB supports).

Indexing runs via the Oban `search_indexing` queue. Its stock configuration starts paused so the search health monitor can control it; verify
that the queue is running before backfilling. A manually managed queue can be enabled with `search_indexing: 10` in the Oban `queues`
configuration instead of a `paused: true` entry.
Monitor that queue until the backfill jobs have drained before evaluating search completeness.

### Rebuilding

Stop Pleroma, or otherwise ensure no search indexing jobs are running, before rebuilding. While ParadeDB is still the configured backend,
run `drop_index`, then `create_index` with the table owner/setup credential. Reapply runtime grants after recreating the table, restart
Pleroma, and begin a new backfill. The DDL tasks print failures without returning a failing process status, so verify the exact `Index
dropped` and `Index created` success messages. Dropping the table removes only derived search data; Pleroma activities remain in the primary
database.

### Rolling back

To retain the ParadeDB table for a later retry, configure `Pleroma.Search.DatabaseSearch` (or the previous external backend) in static config
and restart Pleroma.

Before switching backends, inspect the `search_indexing` queue. Drain or cancel pending backfill jobs, or explicitly accept that they will run
against the newly configured backend after restart.

To remove the table during rollback, first stop Pleroma and run `drop_index` with the table owner/setup credential while ParadeDB is still
configured:

=== "OTP"
    ```sh
    PARADEDB_DATABASE_URL='OWNER_URL' PLEROMA_CTL_RPC_DISABLED=true \
      ./bin/pleroma_ctl search.indexer drop_index
    ```

=== "From Source"
    ```sh
    PARADEDB_DATABASE_URL='OWNER_URL' mix pleroma.search.indexer drop_index
    ```

Verify that the task prints `Index dropped`. Then change the backend in static config and restart. The ParadeDB environment variable may be
removed after removing the ParadeDB stanza or switching to a backend that does not start its Repo. After switching away from ParadeDB, the
generic `drop_index` task targets the newly configured backend and will not remove the ParadeDB table; remove it directly with the database
owner instead.