Configuration

The Configuration page is where you set up everything the daemon needs to function at full capability. Models must be configured for embedding and (optionally) summarization before CI can provide semantic search and session summaries.

Open it from the dashboard sidebar, or navigate to http://localhost:{port}/config.

Models

Models are the engine behind CI’s semantic search and session summaries. You need at least an embedding model for code search. A summarization model is optional but highly recommended for session summaries, titles, and memory extraction.

Providers

OAK supports any local provider with an OpenAI-compatible API:

Provider	Default URL	Notes
Ollama	http://localhost:11434	Most popular choice. Free, easy to set up.
LM Studio	http://localhost:1234	Desktop app with a visual model browser.
Custom	Any URL	Any OpenAI-compatible endpoint (vLLM, llama.cpp, etc.)

Setup steps:

1. Select your provider or enter a custom URL
2. The UI fetches available models from the provider
3. Select a model
4. Click Test & Detect — OAK auto-detects dimensions and context window
5. Save the configuration

Embedding Models

Embedding models convert code and text into vectors for semantic search. You need one running before CI can index your codebase.

Recommended models for Ollama:

Model	Dimensions	Context	Size	Pull Command
nomic-embed-text	768	8K	~270 MB	ollama pull nomic-embed-text
bge-m3	1024	8K	~1.2 GB	ollama pull bge-m3
nomic-embed-code	768	8K	~270 MB	ollama pull nomic-embed-code

Which embedding model should I use?

Start with nomic-embed-text — it’s small, fast, and works well for most codebases. If you need better multilingual support or higher-quality results and have the RAM, try bge-m3. For code-heavy projects, nomic-embed-code is trained specifically on code.

For LM Studio: Search the Discover tab for nomic-embed-text-v1.5 or bge-m3 and download.

For the full list of models OAK recognizes, run:

oak ci config --list-models

Summarization Models

Summarization uses a general-purpose LLM (not an embedding model) to generate session summaries, titles, and extract memory observations. This is a separate configuration from embeddings — you can use different providers or models for each.

Recommended models for Ollama:

Model	Resource Level	Context	Size	Pull Command
gemma3:4b	Low (8 GB RAM)	8K	~3 GB	ollama pull gemma3:4b
gpt-oss:20b	Medium (16 GB RAM)	32K	~12 GB	ollama pull gpt-oss:20b
qwen3:8b	Medium (16 GB RAM)	32K	~5 GB	ollama pull qwen3:8b
qwen3-coder:30b	High (32+ GB RAM)	32K	~18 GB	ollama pull qwen3-coder:30b

Which summarization model should I use?

For most machines, gemma3:4b or qwen3:8b are good starting points. Summarization quality scales with model capability — a more powerful model produces better session titles, more accurate summaries, and catches more memory-worthy observations. If you have the hardware, a larger model is worth it. If summaries are slow, drop down a tier.

For LM Studio: Search the Discover tab for any of the models above and download. Make sure to start the local server with your chosen model loaded.

To see what summarization models are available from your provider:

oak ci config --list-sum-models

Context Window Tips

Local models often default to small context windows that limit summarization quality. The context window determines how much of a session the model can “see” at once.

• Ollama: Default num_ctx is typically 2048 tokens. For better summaries, increase it:

  # Create a Modelfile with larger context
  echo 'FROM gemma3:4b
  PARAMETER num_ctx 8192' > Modelfile
  ollama create gemma3-4b-8k -f Modelfile

• LM Studio: Check the model’s context length in the UI settings and increase if needed.

Higher context windows = better summaries, but more memory usage. 8K is a good minimum for summarization. Models like qwen3:8b support up to 32K natively.

You can also auto-detect and set the context window via CLI:

oak ci config --sum-context auto    # Auto-detect from provider
oak ci config --sum-context 8192    # Set explicitly
oak ci config --sum-context show    # Show current setting

External Resources

• Ollama Model Library — Browse all available models
• Ollama Embedding Models — Filter for embedding-specific models
• LM Studio — Desktop app for browsing and running local models
• llama.cpp — High-performance C++ inference engine; use its OpenAI-compatible server as a custom provider
• MTEB Leaderboard — Benchmark for comparing embedding model quality

Backup Settings

Configure automatic backups and related policies from the Teams page or via the configuration file. See Teams — Automatic Backups for the full guide.

Setting	Config Key	Default	Description
Automatic backups	backup.auto_enabled	false	Enable periodic automatic backups
Include activities	backup.include_activities	true	Include the activities table in backups
Backup interval	backup.interval_minutes	30	Minutes between automatic backups (5–1440)
Backup before upgrade	backup.on_upgrade	true	Create a backup before oak upgrade

Backup settings are per-machine (stored in .oak/config.{machine_id}.yaml), except on_upgrade which is project-level (stored in .oak/config.yaml).

# Per-machine backup settings
codebase_intelligence:
  backup:
    auto_enabled: true
    include_activities: true
    interval_minutes: 30

# Project-level backup settings (in .oak/config.yaml)
codebase_intelligence:
  backup:
    on_upgrade: true

Session Quality

Control when background jobs process sessions:

Setting	Description	Default
min_activities	Minimum activity count before a session qualifies for background processing	Varies
stale_timeout	How long (in seconds) an inactive session sits before cleanup	Varies

• Higher min_activities = only substantial sessions get summarized (reduces noise)
• Lower stale_timeout = sessions are cleaned up faster (more responsive, but may cut off sessions that pause briefly)

Logging

Configure log rotation to manage disk usage:

Setting	Description
Max file size	Maximum size of each log file before rotation
Backup count	Number of rotated log files to keep

See the Logs page for details on viewing and filtering logs.

Directory Exclusions

Control which directories are skipped during codebase indexing.

Built-in Exclusions

OAK includes sensible defaults out of the box:

• .git, node_modules, __pycache__, .venv, venv
• dist, build, .next, .nuxt
• And other common build/dependency directories

OAK also respects your project’s .gitignore — anything gitignored is automatically excluded from indexing.

Custom Exclusions

Add your own patterns from the Configuration page. Patterns use glob syntax (e.g., vendor/**, generated/**). Custom exclusions are saved to the daemon’s configuration file.

Reset to Defaults

If you’ve added patterns you no longer need, use the Reset to Defaults button to restore the built-in exclusion list.

Test & Detect

The Test & Detect buttons on the Configuration page let you verify your provider setup:

• Test connection — Verifies the provider URL is reachable and the model exists
• Auto-detect dimensions — Sends a test embedding to determine the model’s vector dimensions
• Discover context window — Probes the model to find its maximum context length

This saves you from having to look up model specifications manually.