Your agent already knows which project this is

LoreDocs released vault prime and workspace-scoped auto-vaults, which let AI coding agents load full project context with a single idempotent call, `vault_open_workspace(path)`, instead of manually pasting schemas and files at the start of each session. The vaults are stored in a local SQLite file that moves with the project directory, and vault prime injects vault metadata, tags and the latest version of each document into the model prompt, backed by FTS5 full-text search and, on the Pro tier, LanceDB vector search using BGE-small embeddings with BM25 ranking and RRF fusion. The company says manual context delivery for a medium-complexity data project can consume several hundred tokens per session, a cost vault prime reduces to a compact pre-indexed payload.

Every session starts with the same question: "What project are we working on?" For a data engineer or an AI practitioner the answer is never one line. It is a schema definition, a handful of transformation scripts, a note about the data catalog conventions, maybe a reference to last week's experiment. Each token spent re-establishing that context is a token that could have been used to explore a model, run a query, or write a transformation. The session-start tax is real, and it compounds across every short-lived session your pipeline spawns. The problem is not the model. It is the way context gets delivered. Traditional approaches require you to manually point the assistant at a directory, upload a file, or run a one-off import script -- steps that repeat for every new workspace and every new session. When you are iterating on a data pipeline that touches dozens of tables, hundreds of lines of configuration, and a set of experiment notebooks, that overhead adds up fast. LoreDocs vault prime and workspace-scoped auto-vaults remove that step entirely. Workspace-scoped auto-vault: one call, one vault, zero configuration vault open workspace path checks whether a vault already exists for the given directory path. If it does, the same vault is returned. If not, a new vault is created and bound to that directory. The operation is idempotent -- you can call it at the start of any script, notebook, or CI job without worrying about duplicate vaults or race conditions. Because the vault lives in the same file hierarchy as your code, the relationship between source files and their knowledge representation is explicit and portable. Everything is stored in a local SQLite file. Copying the project directory moves the knowledge base with it. No cloud dependency, no configuration files to sync, no vault identifier to remember. The before is familiar to anyone who has worked on a multi-project codebase: open a session, paste the schema, describe the conventions, attach the relevant context. The after is a single deterministic call at session start. The vault resolves in milliseconds. The model already has the picture. For teams running many short-lived sessions -- CI pipelines, automated review agents, nightly diagnostics -- the reduction in per-session token usage is straightforward to measure. Manual context delivery for a medium-complexity data project can consume several hundred tokens per session. Vault prime delivers the same information as a compact, pre-indexed payload at a fraction of that cost. Vault prime: full project context in one call Even with an auto-vault, the model still needs to know which pieces of knowledge matter for the current task. Vault prime solves that. A single call injects the full context of the active vault -- metadata, tags, and the most recent version of each document -- into the model's prompt. The model can answer questions about data schemas, experiment results, or pipeline steps without you listing individual files. Vault prime works alongside the built-in FTS5 full-text search. When a query includes domain-specific terms, the search layer retrieves matching documents across all vaults, applies any tag filters, and merges the relevant snippets into the prime payload. The model sees both the broad project outline and the specific evidence that answers the query. The "show me the schema" prompt disappears; the model already has it. For teams on the Pro tier, a hybrid semantic search layer adds LanceDB vector search on top of FTS5, using BGE-small embeddings with BM25 ranking and RRF fusion. Queries that lack exact keyword matches -- "how do we handle late-arriving records?" against a vault full of domain-specific jargon -- resolve correctly because the search layer understands meaning, not just tokens. The index is built once with vault rebuild index and stays current as documents are added. Bringing existing knowledge in LoreDocs does not require starting from scratch. vault import dir points at an Obsidian vault root and traverses subfolders automatically, extracting YAML frontmatter tags and creating corresponding documents. The same path handles Notion exports. For plain text files, vault add doc accepts a file path and ingests content directly. All imports are versioned: every change to a document creates a new history entry, and you can roll back to any prior version with a single command. The versioning system lives alongside the FTS5 index, so you can search across historical revisions as easily as the latest content. For Pro users, auto-discovered document relationships surface related notes, code snippets, or experiment logs without manual linking. What this means for your pipeline A typical session with LoreDocs looks like this: the first line of your script calls vault open workspace with the project directory, which either reuses an existing vault or creates a new one in milliseconds. A subsequent vault prime call injects the full project context so the model already has the picture. If the query includes domain-specific terms, the FTS5 engine adds the most relevant snippets -- all of this resolves before the model sees the first user token. The session-start tax drops to one function call. Context is consistent across every session in that workspace. New team members and fresh containers get the same knowledge base without any manual setup. And because the entire stack runs locally, there are no round-trips to a remote service and no data leaving the machine. The free tier supports up to three vaults, which covers most solo projects. The Pro tier removes the vault limit and adds semantic search and auto-discovered relationships. Explore the full LoreDocs toolset at /tools https://labyrinthanalyticsconsulting.com/tools or get posts like this delivered weekly via Dispatches from the Labyrinth https://labyrinthanalytics.substack.com/subscribe?utm source=website&utm medium=blog&utm campaign=substack subscribe .