@brain-core

Config settings

brain-core's platform settings — embedding, sync and watching, indexing limits, reading, search and listing, vector maintenance, review, discovery and git — and where to change them.

Every tunable brain-core has is a declared platform setting, edited by the owner in Admin → Config (and the embedding ones also in the admin Vector section or during setup). Each setting shows its label, description, default and bounds there; the manifest's configSettings list is the complete set. Limits default generous. How settings are declared and resolved is on package-system config settings.

Embedding

SettingDefaultMeaning
embeddingEnabledtrueOff makes no document or query embedding calls at all; content and metadata are still indexed
embeddingModelId · embeddingDimensionollama:nomic-embed-text · 768The target model and dimension — changing them starts nothing until a rebuild (Sources)
embeddingEndpointsnoneYour own OpenAI-compatible embedding servers (below)
ollamaUrl · openaiBaseUrl · qwenEmbeddingBaseUrlthe providers' public URLsWhere the local and cloud providers are reached
embeddingDailyUsdCap0 (no cap)Stops embedding for the rest of the UTC day once today's priced spend reaches it; indexing continues and semantic search falls back to text search with a warning
vectorKeepPreviousGenerationfalseKeep the old collection after a rebuild as a rollback copy, at double the storage
embeddingRequestTimeoutMs · vectorEmbedMaxRetries · brainOllamaEmbedConcurrency · vectorOllamaKeepAlive—Request timeout, retries, local concurrency, and how long Ollama keeps the model loaded

Supported providers are Ollama (local), OpenAI, Gemini, Qwen, and Voyage. The model list is a registry of explicit records — provider, protocol, default and supported dimensions, credential id. Two models are multimodal — Gemini Embedding 2 (images, PDF pages, audio and video) and Voyage's multimodal model (images and PDF pages); the rest are text-only. Models that tell queries and documents apart are told which one each call is, and many expose MRL (Matryoshka) selectable dimensions: a smaller vector trades a little quality for less storage and faster search.

Any server that speaks the OpenAI /v1/embeddings protocol can be added as an embedding endpoint: its URL, its models with the dimension each returns (and optionally a price), its authentication, and its address class. Its models appear as endpoint:<id>:<model>, and a keyed endpoint's key is a credential. An endpoint whose model is the active one, the target, or a running rebuild's target cannot be removed, disabled or renamed (409 endpoint_in_use) — switch the index away from it first. A public endpoint is address-checked and IP-pinned on every call, like the cloud providers; private is your explicit statement that the address is inside your own network. The two cloud base-URL overrides accept public addresses only — a gateway inside your network is added as a private endpoint instead.

Sync and watching

brainAutoSyncMode (on, off, packages-only) controls the background scheduler; brainSyncPollIntervalMs (60 s), brainSyncReconcileEveryMs (24 h), brainSyncBootReconcile and its delay (30 s) its cadence; brainSyncMaxConcurrent (2 running jobs) and brainSyncConcurrency (files in parallel inside one job) how much runs at once; brainSyncDebounceMs and brainSyncDirtyUriCap the watcher's batching; brainWatchMaxFolders (200 000 folders) and the two watch-recovery delays the OS watch; brainMaxScanFiles (200 000, overridable per source) and brainScanTimeoutMs the walk; brainQdrantWatchIntervalMs (30 s) the vector-database probe and brainQdrantRequestTimeoutMs (300 s, read at start) the bound on every vector-database request, brainQdrantProbeTimeoutMs (5 s, read at start) how long that probe waits before it counts the vector database as unavailable.

Indexing limits

brainSyncMaxFileMb (5), brainSyncMaxIndexTextChars (200 000) and brainSyncMaxAppendFileMb (20) bound what one file costs; brainSyncBinarySniffChars (8 192, 0 off) skips binaries; brainSyncMaxMediaMb (8) and brainSyncMaxPdfPages (20) bound media embedding; vectorChunkTargetChars, vectorChunkOverlapChars, vectorChunkMinThresholdChars and vectorMaxChunksPerFile shape chunking.

Reading and media

brainMaxReadFileMb (100) caps a read; inlineImageMaxMb, inlineAudioMaxMb, inlineVideoMaxMb, inlinePdfMaxMb and brainVisionBatchBudgetMb cap what fs_read hands a model as media; fsReadPdfRender (off) renders a PDF's first page on read; brainPdfTextExcerptBytes and brainReadSnippetChars bound text excerpts.

Search and listing

brainSearchSemanticDefaultLimit (10), brainSearchFilterDefaultLimit (50) and brainGrepDefaultLimit (50) are the default result counts; brainGrepContextConcurrency bounds grep context reads; vectorExactSearchMaxPoints (20 000) is where semantic search switches from an exhaustive comparison to the approximate index. brainListDefaultLimit (500), brainListSummaryBudgetChars and the connector, git, folder-enrichment and scoped-scroll limits bound fs_list; brainCountCacheTtlMs and brainAccessibleCacheTtlMs cache counts and the visible-source set.

Vector store and maintenance

vectorGcIntervalHours (24, 0 off) and vectorGcRetentionHours (24) drive garbage collection; vectorHealthBatchSize the health scans; vectorWriteHotUriThreshold (60 writes an hour), vectorWriteHotUriTopN and vectorWriteMonitorMaxUris the hot-file report; vectorDefaultSegmentNumber, vectorMaxSegmentSizeKb, vectorMaxIndexingThreads and vectorQuantizationMode (off) the collection settings an owner applies from Vector Health; vectorStoreWriteMaxRetries and its delay the write retries.

Review, discovery, git and debug

  • Review: pendingPreviewMaxBytes (262 144), pendingJournalCacheBytes, pendingRepairBatchSize (50) and pendingOrphanProbeConcurrency.
  • Source packages and discovery: the brainSourcePackageReconcile… limits (roots, entries, concurrency) and the brainDiscoveryInventory… cache.
  • Git: brainGitBadgeTimeoutMs, brainGitPanelTimeoutMs, brainGitHeavyTimeoutMs, brainGitDiscoveryMaxDirents, brainGitDiscoveryMaxRepos and brainGitUntrackedCountCap (the per-refresh budget for counting new files' lines).
  • Debug: brainDebug and brainVectorLogConsole.

On this page