BrainTube

smoke test stub

reproduces the 0cbcf01 incident pattern

Full-text search over your personal BrainTube knowledge corpus. Searches across YouTube, Instagram, web, LinkedIn, GitHub, Twitter and more. Returns results ranked by recency with taint warnings.

Get full details for a specific saved item including transcript, description, summary, key takeaways and taint level. Pass YouTube video ID or internal UUID.

List your most recently saved items across all source types. Use to resume a research session or review what was captured lately.

Get your personal corpus statistics: total items saved, breakdown by source type (youtube/instagram/web/etc), taint distribution. Call this before searching to understand what knowledge is available.

Find items semantically similar to a given item using vector similarity. Useful for discovering related concepts, follow-up research, or building knowledge clusters. Requires embeddings — run backfill_embeddings first if results are empty.

Search your corpus filtered to a specific source type. Use when you want results only from "youtube", "instagram", "web", "notion", "linkedin", "twitter", "github", "reddit", "pdf", "note", etc. Combines semantic + keyword fallback.

Semantic search scoped to items saved between two dates. Pass ISO 8601 dates for "after" and "before". Useful for reviewing what you captured during a specific period or project.

List saved bookmarks from your corpus. Filter by read/unread status. Returns title, URL, tags, and read state sorted by bookmarked_at desc.

Bookmark an item, remove a bookmark, or toggle its read/unread state. Actions: bookmark | unbookmark | mark_read | mark_unread.

Connect your Readwise account to BrainTube by saving your Readwise API token. Required before sync_readwise can run. Get your token at readwise.io/access_token.

Import highlights from your Readwise library into your BrainTube corpus. Use mode=incremental (default) to fetch only new highlights, or mode=full to re-import everything. Requires connect_readwise first.

Ingest a single Notion page into your BrainTube corpus. Accepts a full Notion URL or raw page UUID. Extracts title + body text, upserts to items table, and immediately generates an embedding. Requires set_notion_api_key first.

Write a note or AI-generated synthesis back to a specific item in your corpus. No write_token needed — your JWT proves ownership. Ownership is verified server-side.

Retrieve your most recently saved Claude and ChatGPT conversations. Useful for resuming context from a previous session or reviewing past AI-assisted work.

Analyse your corpus to build a knowledge expertise profile. Tags are classified as expert (>50 items), intermediate (20-50), surface (5-20), or blind spots (<5). Also returns dominant source types and recent 7-day focus.

One-shot session bootstrap: combines expertise profile, last 5 AI conversations, and corpus stats into a single JSON object. Call this at the start of a session to load full context without multiple round-trips.

Ingest a single piece of content (note, article, document, etc.) into your BrainTube corpus. Deduplicates by source_url first, then by exact title match. Immediately generates an embedding. Use force_new=true to skip dedup and always insert.

Ingest up to 50 items in a single call. Each item is deduplicated (source_url → title fallback), inserted or updated, then embedded in batches of 20. Daily limit: 500 new items per user. Returns { inserted, updated, skipped, errors[] }.

Surface forgotten items from your corpus using weighted randomness — items you\'ve retrieved least often are most likely to appear. Great for spaced repetition and rediscovering old saves.

Ingest all pages from a Notion database into your BrainTube corpus. Processes up to `limit` pages with 350ms delay between each. Requires set_notion_api_key first.

Save your Notion integration API key so ingest_notion_page and ingest_notion_database can access your Notion workspace. Get your key from https://www.notion.so/my-integrations.

Generate and store vector embeddings for all your items that are missing them. Required before semantic search works. Processes in batches of 20 with 200ms delays. Returns { embedded, errors } count.

Generate a personal API key for use with the Obsidian Sync CLI or other integrations. The raw key is shown ONCE — save it immediately. Keys are stored as SHA-256 hashes only.

Search your local Obsidian vault via the Obsidian Local REST API plugin (exposed through Tailscale). Returns matching notes with title, file path, and a text excerpt. Requires OBSIDIAN_BRIDGE_URL and OBSIDIAN_API_KEY set in Railway env vars.

Ask a question to a public BrainTube Brain (a curated knowledge base built from someone\'s corpus). Pass the brain_slug (visible in the Brain\'s URL), your question, and optionally prior chat_history for multi-turn conversations. Returns answer + source citations.

List all Brains you have created. Returns slug, name, description, item count, tier (free/pro), and visibility (public/private). Use the slug with chat_with_brain to query a specific Brain.

Build a knowledge graph around a specific item, showing how it connects to other items in your corpus via knowledge_edges. Returns the center item, connected nodes with metadata, and typed edges with confidence scores. Use depth=1 for direct connections, depth=2-3 for wider neighbourhood exploration.

Run a health check on your knowledge corpus. Returns total items, missing embeddings, missing enrichment, missing tags, orphan items, stale items (90d+), contradictions, overdue reviews, topic gaps, and an overall health score out of 100.

Get a topic-level index of your entire knowledge corpus. Groups items by primary topic and returns item count, synthesis count, average salience, latest save date, and source types per topic — sorted by item count descending. Use to understand which subjects dominate your library.

Export your entire BrainTube knowledge corpus as a ZIP file. Items are exported as Markdown with YAML frontmatter, concept articles go into a wiki/ folder, and an index.md is included. Returns a signed download URL valid for 1 hour.

Invoke the compile-knowledge edge function to generate concept articles from a topic cluster or Brain. Synthesises saved items into structured wiki-style articles with backlinks and knowledge graph edges. Pass either cluster_id or brain_id.

Query compiled concept articles from your knowledge base. Filter by cluster_id, brain_id, or free-text search against title and body. Returns title, slug, word count, and backlink count per article.

Find tags that frequently appear together across your corpus. Returns pairs sorted by co-occurrence count — useful for discovering implicit topic clusters and knowledge relationships.

Find named entities (people, orgs, tools) that frequently co-appear across your corpus. Returns pairs sorted by co-occurrence count — useful for mapping who/what clusters in your knowledge base.

Detect knowledge gaps in your corpus: thin topics (few items), entities without depth, stale high-value items, topics missing concept articles, and unconnected items with no knowledge edges.

Return the items you retrieve most often, ranked by retrieval_count. Surfaces your highest-utility knowledge — the items you keep coming back to.

Generate a CLAUDE.md-compatible knowledge context file for any Claude Code project. Pulls your top-salience items, compiled concept articles, key topics, and top entities into a structured Markdown block you can drop into any repo.

Trigger a salience score recompute across your corpus via the compute_salience_scores RPC. Use this after bulk ingests or to refresh rankings. Returns { updated_count }.

Multi-hop knowledge search: runs adaptive_search then traverses the knowledge graph from the top-3 results up to max_hops deep, surfacing semantically connected items that a flat search would miss. Returns direct_results + graph_connected + total_nodes_explored.

Get a retrieval quality dashboard for your corpus over the past N days. Covers search hit rates, zero-result queries, top search terms, and result relevance signals.

Find the shortest path between two items in your knowledge graph. Traverses knowledge_edges up to max_depth hops and returns the ordered list of item IDs and edge types along the path, or "no path found" if disconnected.

Trigger an on-demand recompute of graph centrality scores across your corpus via compute_centrality_scores RPC. Run after adding new knowledge edges or after compile_knowledge. Returns { updated_count }.

Get the temporal history of knowledge edges between two specific items — when they were connected, edge types over time, confidence changes. Pass item_a and item_b as UUIDs.

Get current security alert status: unacknowledged alerts, recent security events (24h), taint distribution, active canary triggers, active alert suppressions, and 7-day retrieval quality metrics. Admin only.

Acknowledge a security alert by UUID, marking it as resolved. Optionally attach resolution notes. Admin only.

Temporarily suppress a specific alert type for maintenance windows. Duration 1–168 hours (max 7 days). Requires a reason. Admin only.

Get LLM firewall status — analytics (7d), shadow mode state, active rule versions, and adaptive threshold recommendations (30d). Admin only.

Promote a firewall check from shadow mode to enforcement (will block/modify), or demote back to shadow (log only). Available checks: toxicity, topic_boundary, conversation_risk, token_budget, ingress_probe, exfiltration, policy_compliance. Admin only.

Update a firewall threshold value with automatic versioning. Thresholds: conversation_risk_warn, conversation_risk_block, grounding_minimum, grounding_writeback_gate, exfiltration_entity_limit, exfiltration_yesno_ratio. Admin only.

Rollback firewall rules (thresholds, shadow_config, etc.) to a previous version. List available versions first with firewall_rule_history. Admin only.

View version history for a firewall rule type (thresholds, shadow_config, injection_patterns, pii_patterns, toxicity_patterns, homoglyph_map, token_limits). Shows version number, description, changed_by, and active status. Admin only.