Configuration

Project Settings

Project settings are managed from the Settings page in the dashboard. These control how documents are indexed and how search queries are processed.

Embedding

Setting	Default	Description
Embedding model	`text-embedding-3-small`	Model used to embed documents and queries

Chunking

Setting	Default	Description
Chunk size	`1024`	Maximum token count per chunk
Chunk overlap	`200`	Token overlap between adjacent chunks

Retrieval

Setting	Default	Description
Match threshold	`0.3`	Minimum cosine similarity score (0-1)
Max results	`10`	Maximum number of results to return
BM25 search	`off`	Enable hybrid BM25 keyword search alongside vector search
Reranking	`off`	Enable reranking of search results using a cross-encoder model

When BM25 search is enabled, both vector similarity search and BM25 keyword search run in parallel. Results are merged, deduplicated by document, and the higher score is kept. This improves recall for keyword-heavy queries while preserving semantic search quality.

Generation

Setting	Default	Description
AI Provider	—	Select a registered AI provider (configured at the team level)
Generation model	Depends on provider	Model used to generate AI answers
System prompt	Built-in	System prompt for the LLM

AI Providers are registered at the team level from the AI Providers page. Each provider has its own API credentials and supported models. Project settings reference these providers via dropdown — no API keys are stored at the project level.

Search Parameters

These parameters can be passed in the request body for /search and /ai-search to override project defaults per-request:

Parameter	Type	Default	Description
`query`	string	required	The search query
`max_num_results`	number	Project setting	Maximum chunks to return
`score_threshold`	number	Project setting (`0.3`)	Minimum similarity score (0-1)
`include_steps`	boolean	`false`	Include internal step trace in the response
`filters`	object	`null`	Metadata filters (see Filtering)

AI Search Parameters

Additional parameters for /ai-search:

Parameter	Type	Default	Description
`stream`	boolean	`false`	Enable SSE streaming responses
`include_steps`	boolean	`false`	Include internal step trace

Filtering by Metadata

Filter results using document attributes:

{
  "query": "How do I deploy?",
  "filters": {
    "type": "and",
    "filters": [
      { "key": "folder", "value": "guides/" }
    ]
  }
}

Filter Types

Type	Description
`and`	All conditions must match
`or`	Any condition must match

Filter Conditions

Key	Description
`folder`	Document folder path prefix (auto-generated from filename)
Custom keys	Any attribute set during document upload

Document Attributes

Attach custom metadata when uploading documents:

curl -X POST https://brain.aistack.run/api/v1/documents \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@guide.md" \
  -F "filename=docs/guide.md" \
  -F 'attributes={"author":"alice","version":"2.0"}'

These attributes can then be used in search filters.

Upload Filtering

Upload filtering lets you automatically reject files based on glob patterns before they are stored or indexed. This is useful for keeping unwanted files (build artifacts, lock files, environment files, etc.) out of your project.

Default Ignore Patterns

The following patterns are always enforced and cannot be disabled. Files matching any of these are silently skipped:

**/node_modules/**    **/.git/**       **/.svn/**
**/.hg/**             **/.DS_Store     **/Thumbs.db
**/*.lock             **/package-lock.json
**/yarn.lock          **/pnpm-lock.yaml
**/bun.lockb          **/.env          **/.env.*
**/*.log              **/*.min.js      **/*.min.css
**/*.map              **/*.d.ts

Per-Project Patterns

You can configure additional patterns per project in the dashboard under Settings > Upload Filters:

Ignore patterns -- Additional glob patterns to reject. These work alongside the default patterns.
Allow patterns -- If set, only files matching at least one allow pattern are accepted. Leave empty to allow all non-ignored files.

Evaluation Order

1. Check default ignore patterns -> skip if matched
2. Check project ignore patterns -> skip if matched
3. If project allow patterns exist:
     File must match at least one -> skip if not
4. Allow upload

Ignore patterns always take precedence over allow patterns.

Skipped Response

When a file is filtered, the API returns HTTP 200 with:

{
  "skipped": true,
  "reason": "File matches default ignore pattern: **/.env"
}

No document is created, no storage is used, and no indexing is triggered.

Configuration

On this page