Indexes

The AI sees what exists. Always.

Every turn, your agent reads a workspace index (channels, profiles, media, artifacts) and a channel index (threads with summaries). Loading the right context isn't a guess — it's a lookup.

Compare to ChatGPTGet started
Indexes

A table of contents the model actually reads.

Boardbox keeps two indexes — one for the workspace, one per channel. Agents consult them like a librarian’s card catalog: cheap to scan, decisive about what to load, surgical about what they pull into context.

Learn more
Workspace Index

Cross‑channel knowledge map

One source‑of‑truth listing every channel, every profile, and every artifact in your workspace — with stable IDs.

Workspace: Personal

Channels
  • #general0d242d…b521— general discussion
  • #career02fe3f…ae7d— career strategy & growth
  • #health493b74…6936— health, fitness, nutrition, wellness
  • #finance3aefc1…8594— cashflow & runway
  • #spiritual99f32a…7af1— faith, theology, devotional life
Profiles
  • Personal9822840a…20d— identity, work, communication
  • Healthb2e1f7…41c— labs, conditions, supplement stack
  • Faith7ac09e…9d2— tradition, practice, study plan
Channel Index

A thread‑by‑thread summary

Every thread in a channel reduces to a one‑line description. The agent reads the index first, decides what’s relevant, and only then loads the thread itself.

Channel: #health

Active threads
  • T‑0142Calprotectin retest plan — 8‑week window, ties to flare risk.
  • T‑0140Two‑week meal plan around the cutting macros from #nutrition.
  • T‑0137Heart‑rate zone training block, 4×/week, paired with sleep score.
Resolved (last 14d)
  • T‑0135April lab panel interpretation — calprotectin flagged.
  • T‑0131Ankle pain protocol after the Eno River run.
  • T‑0128Caffeine taper — finished, holding at 200mg/day.
Two tiers

Workspace index → channel index → the thing itself.

The workspace index lists every channel, profile, artifact, and media file with a short description. The channel index lists every thread (open and resolved) with a one-line summary. The agent reads both on every turn, then calls load_* to fetch the items it actually needs. The index is cheap; loaded items are full-fidelity.

Workspace index

~Channels · profiles · artifacts · media. One line each. Always loaded.

Channel index

Active threads + resolved threads + thread summaries. Loaded when you're in the channel.

Loaded items

Full content of profiles, artifacts, or media. Loaded on demand via tool call.

RAG / vector search

Embed the query, return k-nearest. Sometimes the right doc isn't in the top-k. You can't tell.

Index lookup

The agent sees the full table of contents and decides what to load by name. You can audit the trace.

Why not vectors

Lookup beats search, when you can afford lookup.

Vector search trades exactness for scale — it’s how you handle a billion documents. A single user’s workspace is not a billion documents. It’s tens to a few hundred items, each with a clear identity. At that scale, a structured list the agent can read directly beats embedding search on accuracy, cost, debuggability, and predictability.

Visible operations

Every load is a trace you can read.

When the agent loads a profile, an artifact, or a media file, the call appears inline in the thread as a small trace line. You see what it pulled, when, and why. No black-box retrieval. If the agent loaded the wrong file, you can ask it to fix that — or just edit your index entry to make the right file more findable.

load_profile·Healthload_artifact·PDF · Lab panel — April 2026load_media·IMG · Knee MRI · 2026-03-12load_thread·"Calprotectin retest" · resolved · 2 weeks ago
Liveness

The index updates as your workspace grows.

Add a profile, the index updates. Resolve a thread, the summary lands in the channel index. Upload a file, the workspace index picks it up with an auto-generated title and description from the vision model. The next agent turn sees the new state — nothing to re-index, nothing to re-embed.

FAQ

Common questions

Is this a vector database?
No. Indexes are structured Markdown tables of contents — lists of channels, profiles, artifacts, media, and thread summaries with their IDs and short descriptions. The agent reads them deterministically every turn. No embedding search, no retrieval ranking, no cosine-similarity surprises.
How big can a workspace get before the index becomes a problem?
Indexes use one-line summaries per item, so a workspace with hundreds of artifacts and thousands of threads still fits in a few thousand tokens. We compress thread bodies into summaries automatically; only the index lives in the frontload, not the full content.
Is the index something I can see?
Conceptually, yes — the index is what the agent reads every turn. The dedicated viewer for it is on the roadmap. Today, you can see the same content surfaced through the items it lists (channels, profiles, artifacts, media) directly in the Boardbox app.
How is this different from RAG?
RAG = retrieve-augmented generation: vectorize chunks, embed the query, return nearest neighbors. Indexes are the opposite: the agent sees the whole table of contents, decides which item to load by name, then loads that item in full. Cheaper, more predictable, and you can audit exactly what the model saw.
What happens if the AI loads the wrong thing?
Every tool call (load_profile, load_artifact, load_media) is visible inline in the thread as a trace. If the agent loaded the wrong file, you see it — and you can ask it to load the right one. No silent retrieval errors.
See also

Related features

Profiles

Profiles appear in the index so the agent always knows which exist and can decide which to load this turn.

Learn more

Media

Every uploaded file is auto-titled and listed in the workspace index — the agent can recall it weeks later.

Learn more

Channels

Channels appear in the workspace index with their description, so agents in one channel know what other channels exist.

Learn more

The chatbox was a phase.
Build a workspace.

Boardbox keeps your context, holds its ground, and produces work that lasts.

Get started →