Similar Issues

Overview

Note: /similar_issue is an experimental feature. It works only on GitHub, carries a disproportionately large share of the project's dependency and configuration surface for a single-provider tool, and is therefore excluded from the v1 stability guarantees. No backend is tested against its real driver.

The similar issue tool retrieves the most similar issues to the current issue. It is an issue-scoped command: comment /similar_issue on the issue (served by the GitHub Action; the GitHub App webhook only dispatches comments on pull requests), or run it from the CLI with --issue_url:

/similar_issue

Example usage

Indexing and re-runs

To perform retrieval, the similar_issue tool indexes the repository's issues in the configured vector database. On every backend:

• The first run indexes up to max_issues_to_scan issues (default 500).
• Later runs check the newest max_issues_to_scan issues and index any that are missing, so previously-indexed issues are not re-embedded.
• force_update_dataset = true makes every run re-index the whole repository: LanceDB deletes the repository's rows first, Pinecone and Qdrant upsert over the existing rows.
• skip_comments = true skips issue comments and embeds only the issue title and body.

These keys live under the [pr_similar_issue] section. Each backend section below describes its own configuration and re-run behaviour.

Selecting a Vector Database

Configure your preferred database by setting the vectordb parameter under [pr_similar_issue] in configuration.toml:

[pr_similar_issue]
vectordb = "lancedb"  # options: "pinecone", "lancedb", "qdrant"

Available Options

Choose from the following Vector Databases:

1. LanceDB
2. Pinecone
3. Qdrant

LanceDB Configuration

LanceDB is the default backend (vectordb = "lancedb") and needs no external credentials. The index is stored in a local directory given by the [lancedb] uri key (default ./lancedb). A single table (codium-ai-pr-agent-issues) is shared by every repository indexed into the same directory; rows are tagged with the repository name in the metadata.

As described in Indexing and re-runs, the first run creates the table and indexes up to max_issues_to_scan issues; later runs index any missing issue among the newest max_issues_to_scan; force_update_dataset = true deletes the repository's rows and re-indexes them.

Pinecone Configuration

To use Pinecone with the similar issue tool, add these credentials to .secrets.toml (or set as environment variables):

[pinecone]
api_key = "..."
cloud = "aws"
region = "us-east-1"

The cloud value must be one of aws, gcp or azure, and region must be an availability region offered by that cloud. These parameters can be obtained by registering to Pinecone. Note that the tool uses Pinecone's serverless index API; the former environment setting from the gcp-starter pod tier is no longer supported.

cloud and region are only used when the index does not exist yet and needs to be created. An existing index is opened by name and is never recreated, so moving an existing deployment to the new configuration does not lose the stored vectors.

On re-runs, the first run creates the index and upserts up to max_issues_to_scan issues; later runs index any missing issue among the newest max_issues_to_scan; force_update_dataset = true re-indexes the whole repository.

No backend is tested against its real driver

The similar-issue dependency group is not installed in CI, so the pinecone tests run against a faked module, the qdrant tests never construct a client, and the lancedb tests run against a fake table.

Default vector database

vectordb defaults to lancedb, which works with no external credentials. To use qdrant or pinecone, set vectordb = "qdrant" or vectordb = "pinecone" under [pr_similar_issue].

Qdrant Configuration

To use Qdrant with the similar issue tool, add these credentials to .secrets.toml (or set as environment variables):

[qdrant]
url = "https://YOUR-QDRANT-URL" # e.g., https://xxxxxxxx-xxxxxxxx.eu-central-1-0.aws.cloud.qdrant.io
api_key = "..."

Then select Qdrant in configuration.toml:

[pr_similar_issue]
vectordb = "qdrant"

You can get a free managed Qdrant instance from Qdrant Cloud.

api_key must be present even when the server does not enforce authentication (an empty string is accepted): the tool reads both url and api_key and raises when either is absent. A re-index uploads the whole repository in a single request.

Qdrant points are stored in a collection named codium-ai-pr-agent-issues-v2, derived by appending a -v2 suffix to the shared index name (codium-ai-pr-agent-issues). The suffix is an implementation detail of the Qdrant backend only; pinecone and lancedb use the unsuffixed name.

On re-runs, the first run creates the collection and stores up to max_issues_to_scan issues; later runs index any missing issue among the newest max_issues_to_scan; force_update_dataset = true re-indexes the whole repository.

Upgrading an index created before the point-id fix

Earlier versions derived the point id from the issue id alone, so the same issue number collided across repositories. The id is now seeded with the repository name, which means points written by an earlier version are never rewritten or deleted - they still carry a matching metadata.repo payload, so they stay queryable and can surface alongside their replacements.

The -v2 collection suffix sidesteps this: the new index is written to codium-ai-pr-agent-issues-v2, leaving the pre-existing codium-ai-pr-agent-issues collection untouched. Nothing is deleted, and the first run after the upgrade re-indexes the repository into the new collection. Once you are satisfied with the results, you can delete the old codium-ai-pr-agent-issues collection from Qdrant by hand to reclaim the storage.

How to use

• Install the tool's extra dependencies (vector databases and datasets), which a bare uv sync does not include: uv sync --group similar-issue

• To invoke the 'similar issue' tool from CLI, run: uv run pr-agent --issue_url=... similar_issue

• To invoke the 'similar' issue tool via online usage, comment on an issue: /similar_issue

• You can also enable the 'similar issue' tool to run automatically when a new issue is opened, by adding it to the pr_commands list in the github_app section