VibeCBS collects project nicknames in advance: the agent finds 2.8 times more documents
One project can live under three names: "Menu for the Week" in the tracker, "Weekly Food Subscription" in the wiki, and "Dinnerator" in the chat. An agent that only knows one name will answer "we ship on Friday" and never see the chat where the launch was already postponed. VibeCBS collects such aliases in advance. Across 300 epics (large projects), search returned 13,178 documents instead of 4,741, which is 2.8 times more. Aliases should be looked up before the first question arrives.
AI-processed from Habr AI; edited by Hamidun News
One and the same project almost always lives under several names inside a company, and the search agent usually knows only one of them. Imagine a fictional project called "Menu for the Week": starting Friday, the company wants to sell dinners for seven days at once. You open an epic (a large task that bundles smaller ones) in the tracker.
All the tasks are closed, and the last comment says "discussed in the chat." Which chat exactly is unclear. A developer is surprised: "But that's in Dinnerator."
Meanwhile, the requirements page in Confluence is titled "Weekly Food Subscription." Three names for one project: these are aliases. The team knows them all, but the agent does not.
The developer of the VibeCBS platform used production data to show how much this narrows the search: across 300 large epics, the number of documents found rose from 4,741 to 13,178, that is, 2.8 times, once aliases became part of the query.
Why the agent confidently answers wrong
If the agent has only the name from the tracker, it will find the closed tasks and the plan with the Friday date. It will have real links, and it will answer: "Yes, we ship on Friday." The answer looks reliable because it is backed by sources. But the product manager's decision to postpone the launch sits in a Telegram thread where the project is called "Dinnerator."
The author of VibeCBS, a platform for working with corporate data, shows how a person reaches that word. First they take the epic number MENU-42 and search for it in Confluence texts. They find requirements under a different title. Following the page address, they find a developer's reply in the chat, and in the question to that developer they see the word "dinnerator." Searching that word in Slack turns up a tester's report, and in Telegram, a question, "Are we keeping Dinnerator in the email?" and the product manager's answer: "We're moving the launch to next week." The old name gave yesterday's "we're sending," the nickname gave today's "we're not sending."
When a word becomes a project name
Not every word works as an alias. If you search just for "subscription," a neighboring chat will turn up "fixed the subscription," but that has nothing to do with weekly orders. That is why VibeCBS moves alias search into a separate stage, and it works in two steps.
A word can be tried in search one utterance at a time: the model proposes the next query and shows where it got the word. The query goes into the project's queue, the code runs the search, filters out already-seen records by their identifiers, and passes only new ones to the model.
Recording a word on the project card is harder. It requires independent confirmations. The model rates on a scale from 0 to 1 how confidently each one links the word to the project, and the code checks whether there is enough evidence.
Three confirmations scoring above 0.8, or two if both scores are no lower than 0.95, are enough.
The thresholds are configurable. Independence matters here: a link to the same tester's report can be forwarded to ten chats, but the report is still one. For "Dinnerator," there are two confirmations: the developer's reply with a link to the requirements and the tester's report on the task from MENU-42.
If the model rates each at least 0.95, the name goes onto the card together with links to the sources.
Production numbers and token economics
On November 21, the author took 300 large epics from 2026 from VibeCBS production. Aliases were found for 217 of them. For each epic, they counted how many distinct documents the search returned: a document can be either a task or a requirements page. Using only epic names, the total came to 4,741 documents. With confirmed aliases from the cards, it came to 13,178. The results rose almost threefold (more precisely, 2.8 times).
A caveat on the method: within an epic, documents are counted without duplicates, but if one page discusses two epics, it lands in the total twice. Moreover, a rise in the number of documents does not equal a rise in useful answers. This is a measure of coverage, not precision.
The main engineering idea is not to make the first user wait. You could have the first agent work things out and save the names for the next ones, but then the person would wait while the agent crawls through tasks and chats. Instead, aliases are searched during bootstrap, meaning the first pass over the entire corporate corpus, and then updated as new messages and task changes arrive. This kind of upfront work is called precomputation.
Cost was measured on one average epic. The code made 76 requests to sources to gather material for the card, and on a repeat run it saw the collection was complete and made no new ones. Processing the material for one card took 7 model calls and about 250 thousand tokens. Paying that price every time someone opens a new chat would be wasteful. Plus, a shared card gives every agent the same result: both the person asking about the launch and a colleague in marketing who needs an email.
What to do in practice
If you are building a corporate RAG system or an assistant agent, check how it finds projects by something other than their official name. Collect aliases in advance, require several independent sources, and store the links to them next to the name. Do not count one and the same forwarded document as multiple confirmations.
This scheme has a limitation too. Even with a card, you need to check how far the correspondence extends: before sending an email, the author still verifies what the product manager wrote last. If the search only has data through Thursday, you will have to open the chat itself or honestly tell your colleague which day your information runs through.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Sakana AI Opens Applied Research Lab for Market and Economic Modeling
- Nov 21, 2026Meta Is Reportedly Preparing the $300 Muse Charm AI Tamagotchi With a Year of Verizon 5G
- Nov 21, 2026Dreamteam Raises $53M for a CRM That Fills Itself from Email and Calls
Key Questions Answered
Why the agent confidently answers wrong?
If the agent has only the name from the tracker, it will find the closed tasks and the plan with the Friday date. It will have real links, and it will answer: "Yes, we ship on Friday." The answer looks reliable because it is backed by sources. But the product manager's decision to postpone the launch sits in a Telegram thread where the project is called "Dinnerator." The author of VibeCBS, a platform for working with corporate data, shows how a person reaches that word.
When a word becomes a project name?
Not every word works as an alias. If you search just for "subscription," a neighboring chat will turn up "fixed the subscription," but that has nothing to do with weekly orders. That is why VibeCBS moves alias search into a separate stage, and it works in two steps. A word can be tried in search one utterance at a time: the model proposes the next query and shows where it got the word.
What are the key details on Production numbers and token economics?
On November 21, the author took 300 large epics from 2026 from VibeCBS production. Aliases were found for 217 of them. For each epic, they counted how many distinct documents the search returned: a document can be either a task or a requirements page. Using only epic names, the total came to 4,741 documents. With confirmed aliases from the cards, it came to 13,178. The results rose almost threefold (more precisely, 2.8 times).
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.