Creates a new Qdrant collection using a named preset (dense, hybrid, or multitenant), with named vectors and payload indexes configured correctly by default. Use this once per new document set that needs its own collection — not for adding documents to an existing collection (use ingest_* for that).

Lists every Qdrant collection managed by this server, with basic size info. Use this to see what RAG collections already exist before creating a new one or picking which to search.

Returns full detail on one collection: vector schema, payload indexes, size, and optimization status. Use this to inspect a specific collection's configuration, e.g. before deciding whether hybrid search or a given filter is available.

Permanently deletes a Qdrant collection and all its data. Destructive — requires confirm_name to exactly match name, and is blocked entirely when the server is in read-only mode. Use only when the user has explicitly confirmed they want a collection gone.

Points an alias at a collection, enabling zero-downtime blue/green re-indexing — clients keep querying the alias while the underlying collection is swapped out. Use this when republishing a freshly re-indexed collection without downtime.

Creates a payload index on a collection field so it can be used as a fast search filter (e.g. date, author, type). Use this when the wizard's filter question or a user request names a field that needs to be filterable.

Creates a point-in-time backup (snapshot) of a collection. Use this before a risky operation (bulk re-ingest, schema change) or on a routine backup cadence.

Restores a collection from a previously created snapshot, overwriting its current contents. Destructive — requires confirm_name to exactly match collection, and is blocked in read-only mode. Use only when the user has explicitly confirmed the overwrite.

Ingests a piece of raw text directly, with optional metadata, without needing a source file. Use this for notes, pasted content, or anything the user dictates in chat rather than pointing at a file — the 'semantic memory' use case, done properly.

Ingests a single local file (PDF, DOCX, XLSX, PPTX, MD, HTML, CSV, or TXT) into a collection, with format-specific quality processing. Use this for one file at a time; use ingest_directory for a whole folder.

Recursively ingests every supported file under a local directory, with optional glob include/exclude patterns, as a background-trackable job. Use this for bulk ingestion of a folder; poll job_status for progress.

Fetches a web page and ingests its main content, stripped of navigation/cookie-banners/footers. Use this for documentation pages, articles, or any URL-addressable content the user wants in the RAG.

Checks the progress of a background ingestion job started by ingest_directory. Use this to poll a long-running bulk ingest rather than assuming it finished — a job is only 'completed' once its processed/failed counters reconcile with what was expected.

Lists ingested source documents in a collection, one entry per original file/text/URL (not per chunk). Use this to see what's already in a collection before ingesting more, or to find a doc_id to pass to document_delete or get_document.

Deletes every chunk belonging to one source document from a collection, without touching the rest. Destructive — requires confirm_doc_id to exactly match doc_id, and is blocked in read-only mode. Use only when the user has explicitly confirmed the deletion.

Plain dense semantic search over a collection, with optional payload filters. Use this for a fast, simple similarity search when hybrid fusion or reranking isn't needed — for most RAG questions, prefer get_context or search_hybrid instead.

Dense + sparse hybrid search with native RRF fusion — the recommended default search tool for most queries, since it handles both semantic meaning and exact keyword/ID matches well.

Hybrid search followed by cross-encoder reranking over the top results, for maximum precision at the cost of extra latency. Use this when result quality matters more than speed, e.g. for a final answer rather than an exploratory search.

Runs several query reformulations (supplied by the calling LLM) and fuses their hybrid results into one ranking. Use this when a single query phrasing might miss relevant chunks — e.g. ambiguous or broad questions.

Finds points similar to a given point by ID, using Qdrant's native similarity API. Use this to explore 'more like this' starting from a specific chunk the user is looking at.

Recommends points using positive and negative example point IDs, via Qdrant's native recommendation API. Use this when the user can point at examples of what they want more or less of, rather than phrasing a text query.

The flagship retrieval tool: runs hybrid search, reranks, applies MMR for diversity, trims to a token budget, and returns a formatted context block with numbered citations — ready for the calling LLM to answer from directly. Prefer this over raw search_* tools whenever the goal is to answer a question, not just to inspect search results.

Fetches the chunks immediately before and after a given result within the same document, for continuity. Use this when get_context's answer references a chunk that seems to be cut off mid-thought.

Retrieves the full original source document behind a citation, or a specific page/section range of it. Use this when the user wants more context than a single cited chunk provides.

Starts a new conversational setup session and returns the first question, with options and a recommended default. Use this once, at the very beginning of guiding a user through configuring a new RAG — never mid-conversation once a session is already in progress.

Records the user's answer to the current wizard question, validates it live (e.g. can Qdrant be reached, does an API key work), and returns either the next question or a null next step once all 8 questions are answered.

Executes the fully agreed wizard plan: creates the collection, indexes, and profile, ingests one example document, and runs a smoke-test search. Use this only after setup_answer has returned all 8 questions answered — never before the session is complete.

Lists saved configuration profiles (e.g. demo, work, project X). Use this to see what profiles already exist before creating a new one or switching.

Activates a saved profile, switching which collection and embedding configuration subsequent tool calls use. Use this to switch between separate RAG setups, e.g. from 'demo' to 'work'.

Reports Qdrant connectivity, which embedding model is loaded, the reranker's live status, server version, and the active transport. Use this first when something seems broken, or to confirm the server is reachable and correctly configured.

Reports usage stats across all collections (or one, if named): point counts, document counts, disk size, and distribution by source type. Use this to answer 'how much do I have in my RAG' style questions.

Before ingesting, estimates how many chunks a file or directory will produce, the storage it will use, and the embedding API cost if a paid provider is configured. Use this to preview a large ingest before committing to it.

Returns the effective configuration of the currently active profile, with secrets (API keys, bearer tokens) masked. Use this to confirm what's actually configured without ever exposing credentials.