This page defines codebase-specific terms, jargon, and abbreviations used throughout Vault Cortex. It serves as a technical reference for onboarding engineers to understand the implementation details and naming conventions used in the system.
The open standard that enables AI models to interact with external data sources and tools. Vault Cortex implements the MCP server specification to expose Obsidian vaults.
A specific MCP transport type that uses HTTP with chunked encoding to allow long-running tool outputs to be streamed to the client. Vault Cortex implements this via an Express-based router.
A scoring model by Glama used to evaluate the effectiveness of tool descriptions for LLMs. Vault Cortex tool descriptions are optimized with specific "Example", "When to use", "Returns", and "Errors" sections to maximize this score.
The SQLite Full-Text Search extension (version 5) used to index note content. It supports prefix matching and BM25 ranking.
A SQLite extension that adds vector search capabilities. It is used to store and query embeddings in the note_chunks_vec table.
An algorithm that merges two independently ranked result lists (FTS keyword + vector semantic) into a single relevance score. It uses the formula 1 / (k + rank) and adds top-rank bonuses.
A secondary ranking stage that uses a model (ms-marco-MiniLM-L-6-v2) to rescore the top candidates from the hybrid search by looking at the interaction between the query and the document text.
A technique used to combine reranker scores with RRF scores using rank-dependent weights (e.g., 75/25 at top ranks, 40/60 at lower ranks).
A performance optimization that uses SHA-256 hashes of note chunks to skip re-embedding unchanged content during index updates.
The strategy of splitting notes into smaller segments for vector search based on Markdown headings, ensuring that each chunk maintains the context of its parent title.
The default local ONNX embedding model used to generate 384-dimensional vectors for note chunks.
The specific cross-encoder model used for reranking search results to improve semantic precision.
The SQLite journaling mode used to allow concurrent readers and a single writer without blocking.
The ranking algorithm used by SQLite FTS5 to calculate relevance scores based on term frequency and inverse document frequency.
A specific SQLite table that tracks non-markdown assets (images, PDFs, .canvas) to ensure wikilinks to these files resolve correctly.
The standard Obsidian link format: [[Note Name]]. The obsidian-markdown/links.ts module parses and resolves these.
The YAML block at the top of a Markdown file containing metadata.
The first block in a note (after frontmatter) if it is an Obsidian callout. Indexed as a summary of the note's purpose.
A file path starting from the vault root, including the extension (e.g., Folder/Note.md).
A note that has no incoming links from other notes in the vault.
An incoming link to a note from another note in the vault.
A link to a daily note that does not yet exist. These are tracked so they don't count as broken links if the target matches the daily note format.
A specialized indexing system that recognizes tasks within Obsidian Kanban boards, preserving lane information (headings) and card order.
The metadata format used by the Obsidian Tasks plugin (e.g., 📅 2024-07-10 for due dates).
The metadata format used by the Dataview plugin (e.g., [due:: 2024-07-10]).
The open format for Obsidian Canvas files. Vault Cortex includes a linearizer to turn these spatial files into readable text.
The "About Me/" system for storing user preferences and principles. Controlled by the MEMORY_ENABLED flag.
The module responsible for parsing dated bullets in memory files into granular entries for the vector index.
A frontmatter convention (e.g., entry-policy: living) used to define how memory entries in a specific file should be treated (e.g., pruned or preserved).
A leading callout in a memory file defining the file's purpose and contents.
The convention for memory entries: - **YYYY-MM-DD**: Content.
A safety mechanism in the memory store that prevents deleting large amounts of content in a single operation to avoid accidental data loss.
An operation that modifies a specific section of a note identified by a heading without rewriting the entire file.
A file write operation that ensures the file is either fully written or not written at all, typically using a temp-then-rename strategy.
A security extension to the OAuth 2.0 authorization code flow used to prevent authorization code injection.
A sorting strategy for tasks that falls back through multiple date fields (due -> scheduled -> start) to provide a logical timeline.
Logic in vault_update_task that identifies the "Done" lane on a Kanban board by looking for **Complete** markers or headings named "Done".
The layer that handles reading and listing non-markdown files, including image fitting and PDF rendering.
The logic that downscales and recompresses images to fit within a specific byte limit for MCP transport.
The convention where operation functions take exactly two arguments: a parameters object and a logger instance.
A wrapper for tool registrations to handle errors and log context consistently.
Folders blocked from destructive operations (e.g., About Me/, Daily Notes).
SST v4 is the IaC framework used to deploy to AWS Lightsail containers.
A path-aware AWS Lambda that validates static tokens or JWTs for /mcp routes while allowing OAuth routes to pass through.
The CLI tool used to run Obsidian Sync in a headless environment.
A process supervisor used in the Docker container to manage the lifecycle of the MCP server and Obsidian Sync.
Environment variables controlling the opt-in/opt-out status of the memory layer, vector search, and reranking stages.
A flag that enables polling for the file watcher and exclusive-write strategies on Windows.
Refresh this wiki
This wiki was recently refreshed. Please wait 6 days to refresh again.