Precursor IQ

Find anything in your workspace by meaning, not exact wording, and ask questions that come back with cited answers. The same engine powers the IQ section, the ⌘K palette and the retrieve / ask MCP tools. It follows the WorkIQ pattern (retrieval returns ranked passages plus a grounding block with [^n] citations), applied to Precursor's own content instead of Microsoft 365.

⌘K ranked by Precursor IQ. The words can come in any order and in any field, and title hits float to the top.

What gets indexed

Precursor IQ keeps a derived index of passages of up to about 1,200 characters, split on paragraph boundaries with a small overlap:

Source	Passages from	Gated over MCP by
Topics	title and description	topics
Topic briefs	the topic summary	topics
Topic conversations	user and assistant messages, plus text extracted from attachments (PDF, DOCX, PPTX, text)	messages
Chats	title, description, messages, attachments	chats
Agents	title, task prompt, final answer	agents
Live sessions	notes, summary, transcript (with speaker names), insights	live
Memory	long-term memory entries	memory

Tool output and system messages aren't indexed. Archived items stay in the index but are hidden from results, so un-archiving brings them back with no re-index.

Staying current

Every write to an indexed row queues it for re-indexing in the same transaction as the write. A background indexer processes the queue every few seconds (PRECURSOR_IQ_INDEX_POLL_SECONDS), and every query first processes a small batch itself, so results include what you just wrote. An hourly check (PRECURSOR_IQ_RECONCILE_POLL_SECONDS) catches bulk edits and deletes that bypass the queue. On first start, Precursor queues everything once and builds the index in the background.

Settings → MCP servers → Precursor IQ shows the passage count, the queue depth and which full-text engine is active, and has a Rebuild index button.

How ranking works

Precursor IQ builds two ranked lists and merges them with Reciprocal Rank Fusion:

• Full-text: BM25 over the index, using SQLite FTS5 (accent-insensitive, so reunion matches réunion) or Postgres tsvector. Each query word also matches as a prefix, and a hit on a title counts more than a hit in the body. On a SQLite build without FTS5, it falls back to a plain substring scan.
• Semantic (optional): cosine similarity between embedding vectors, so paraphrases match too. See Semantic matching.

Titles that contain every query word get a boost, and recent items get a small one. Results then keep only the best passage per item, and at most three items per topic, chat or session, so one long thread can't fill the list.

The IQ section

IQ in the section rail (or /iq) is a single question box. Press ⌘⇧K (Ctrl+Shift+K) from anywhere in Precursor, even mid-typing or with ⌘K open, to jump straight to it. Text you had selected (a line of a message, a meeting note) is pre-filled as the question, ready to edit before you press Enter. Ask something in plain words, or pick one of the example questions:

1. Matching content appears first, within a moment. Each row opens the real topic, chat, agent or live session, and items whose title matches your question also show as Go to shortcuts above the answer.
2. The answer follows, written by your model from those same numbered passages. Click a citation such as [3] to open its source; the ones the answer relies on are marked cited in the list.

The IQ section: ask once, then open the sources behind the answer.

It's a way to find things, not a conversation: questions and answers aren't saved, and asking again replaces the previous answer. Memory entries can be sources, but they have no page to open.

Ask from ⌘K

In ⌘K, type a question and press Tab (or pick Ask Precursor). The top passages go to your model, which writes a short answer that cites them as [n]. The numbered sources are listed under the answer; press Enter on one to open it. If nothing matches, Precursor says so and doesn't call the model.

Ask mode: a cited answer grounded only on your own content, with its numbered sources.

The answer uses your chat model unless you pick another under Answer model. It's a single model call, logged in usage stats as /iq-ask.

Over MCP

The built-in precursor MCP server exposes two tools, each behind its own toggle under Settings → MCP servers → Precursor capabilities. Both are off by default:

Section	Tool	What it returns
IQ retrieve (iq)	retrieve(query, sources?, limit?)	markdown (numbered excerpts to ground on and cite as [^n]) and hits. Each hit carries its citation id, section, entity_id, ref, topic path, in-app url, matched field, snippet, excerpt and the accessor tool that reads the full item.
IQ ask (iq_ask)	ask(question, sources?)	answer with [^n] citations, the citations it used and every sources passage it was given. Makes one model call.

IQ never discloses more than the per-section tools would: every hit is gated by its own section (see the table above). With only iq and topics on, a caller sees topic titles, descriptions and briefs, but not topic messages (messages) or chats (chats). sources narrows the search further, e.g. ["live", "memory"].

Prefer retrieve for questions

search stays a literal substring match: fast and predictable when you know the exact words. retrieve is the tool for natural-language lookups.

Semantic matching (embeddings)

Embeddings are off by default because each indexed passage is sent once to your provider's /embeddings endpoint, which uses quota. To turn them on, tick Semantic matching in Settings → MCP servers → Precursor IQ.

• Model: defaults to text-embedding-3-small, requested at 256 dimensions. On Azure AI Foundry, enter your embeddings deployment name. Models that can't shorten their vectors keep their native size.
• Providers: GitHub Copilot, the OpenAI-compatible providers and Azure AI Foundry. With a provider that has no embeddings endpoint, retrieval stays full-text only and the card says so.
• Storage: vectors are stored next to their passages and compared in-process, with no vector database or extra dependency. Changing the provider or the model re-embeds everything in the background. Embedding calls appear in usage stats as /iq-embeddings.

Configuration

Setting	Where	Default
PRECURSOR_IQ_ENABLED	env	true. When false, ⌘K falls back to substring search and /api/iq/* returns 404.
PRECURSOR_IQ_INDEX_POLL_SECONDS	env	20
PRECURSOR_IQ_RECONCILE_POLL_SECONDS	env	3600
Semantic matching / embeddings model / answer model	Settings → MCP servers → Precursor IQ	off / text-embedding-3-small / chat model
IQ retrieve / IQ ask exposure	Settings → MCP servers → Precursor capabilities	off / off

The background indexer only runs when PRECURSOR_SCHEDULER_ENABLED is on, like Precursor's other background jobs. Without it, the index is still updated on each query, but embeddings aren't computed.