An agent that guesses at policy is a liability with good grammar. Generic RAG bolted onto a chatbot retrieves the wrong clause from a stale PDF and states it with confidence. Grounding here is first-class: workers retrievebefore they answer, retrieval runs semantic and full-text search fused together, the corpus is read-only to every run — and when your documents change, the answers change with them.
In a pack, knowledge bases arrive stocked: sources indexed and wired to the agents that need them.
See Solution Packs →
The failure isn't that ungrounded models are wrong often — it's that they're wrong fluently. Invented policy, a stale price, the wrong clause of the right document, delivered in the same confident register as the truth. Nobody catches it until it's been said to a customer.
So grounding here isn't a bolt-on. Retrieval runs before the answer, searches by meaning and by exact term, and the answer draws on passages you can check against the source. The corpus is read-only to every run, deduplicated as it's maintained, and honest about deletion — what it holds is exactly what your workers can say.
Each is the same mechanism — retrieve first, answer from the retrieved text — applied where a wrong answer costs real money.
A customer asks about a refund past the 30-day window. An ungrounded model improvises something plausible. A grounded worker retrieves the actual clause and answers from it.
Vector search finds passages about refund timing; full-text search pins the exact terms — '30 days', the policy section number. RRF fuses the rankings.
The worker answers from the refund-policy manual's own words, citable to the passage — not from what refund policies usually say.
All support workers on the base ground identically, so the policy is stated once, correctly, everywhere.
An underwriting team's screening agent must apply the guidelines as written — not a model's recollection of how underwriting generally works.
For each application, it retrieves the passages governing that case type — exposure limits, exclusions, referral conditions — by meaning and by exact term.
The output points to the retrieved guideline passages, so a reviewer checks the source, not the model's reasoning from memory.
The base is read-only at run time; when guidelines change, the team updates the corpus and re-syncs — every screening after that uses the new text.
Your product docs change weekly. The workers answering from them should track the changes without anyone redeploying or editing an agent.
Docs live in the base as files, a crawled site or a synced wiki. Workers reference the base, not any particular version of a file.
Content hashing skips unchanged documents; changed pages re-embed. No duplicate chunks, no full rebuilds.
Every attached worker retrieves from the refreshed index on its next question. Zero worker config touched.
Last year's pricing sheet is obsolete. In a typical RAG stack its embeddings keep answering. Here, deleting it deletes its voice.
The old document is deleted from the base by someone with access — a deliberate, human act, not a run's side effect.
Deletion propagates to the vector index: the document's chunks are removed from retrieval, not just its listing.
No worker can retrieve the retired sheet again. What the corpus holds is exactly what workers can cite.
One team curates the authoritative base. Every worker in the workspace grounds on it — without the curators losing control of it.
Files, sites and synced sources in a single corpus, deduplicated as it's maintained.
Agents, teams and employees attach the same base; each retrieves from it at run time.
Every attached worker searches and cites. None can modify it. Curation stays with the curators.
Bring your knowledge in once; every agent, team and employee retrieves from it at run time — and none of them can write to it.
Six properties that separate a grounding layer from RAG bolted onto a chatbot.
An ungrounded model answers policy questions from its training data — fluently, confidently, and sometimes wrong. Attach a knowledge base and the worker searches it at run time first: the passages that actually match the question are retrieved and the reply is drafted from them, with the retrieved material available as evidence. The answer stops being an assertion and becomes a claim with a source.
Pure vector search is famous for a specific failure: it retrieves what's similar and misses what's exact — the SKU, the clause number, the error code. Retrieval here runs two searches in parallel: dense vector search for meaning, and full-text search for the literal terms. The two candidate lists are fused with reciprocal-rank fusion, so a chunk that both searches agree on outranks a chunk only one found.
Point a base at what you already have: pasted text, uploaded files (PDF, DOC, TXT, MD), a crawled website, or synced sources from Google Drive and Notion. Everything is chunked and embedded automatically — there is no pipeline to build and no index to tune. Mix source types in one base and scattered institutional knowledge becomes one searchable thing.
The quiet way RAG rots: every re-upload and re-sync adds another copy, and retrieval starts surfacing three versions of the same paragraph. Ingestion here fingerprints content by hash — an unchanged document is recognized and skipped, and a re-sync re-processes only what actually changed. The index stays one-copy-per-truth instead of accreting duplicates.
In a bolted-on RAG stack, deleting a file often deletes the file — and leaves its embedded chunks answering questions for months. Here deletion propagates: removing a source removes its vectors from the index, so a retired policy can't be retrieved and quoted after it's gone. What the corpus says is what your workers can say.
A knowledge base is read-only at run time — workers retrieve and cite; no run can mutate the corpus. That's a governance property, not a limitation: one misbehaving or prompt-injected worker can never poison the reference the rest of the workforce depends on. Retrieved chunks are also among the untrusted content the platform's injection shield can screen before they re-enter a model's context. Data workers should read and write lives in Tables, under per-verb permissions.
A support org grounds its workers in two bases — the product docs and the refund-policy manual. A customer asks for a refund 45 days after purchase. Here's what the system does with that question.
Workers search the base before answering; replies draw on retrieved passages, traceable to their source.
Vector and full-text search run in parallel and fuse with reciprocal-rank fusion — meaning and exact terms both surface.
Files, crawled websites, pasted text, Drive and Notion syncs — chunked and embedded automatically, mixed in one base.
Content is fingerprinted; unchanged documents are skipped on re-sync, so the index never accretes duplicates.
Removing a source removes its vectors — a retired document can't be retrieved into an answer.
No run can write the base. Updates are deliberate, human acts; the reference can't be poisoned mid-run.
Update the corpus and re-sync; every attached worker answers from the new material with no config change.
One base attaches to many workers — agents, teams, employees — so the same question gets the same grounded answer.
End-to-end procurement operations solution pack with vendor registry, purchase request review, contract path assessment, embedded review team, and AI employee coordination.
Complete operating system for a Chartered Accountant practice. Tracks clients, engagements, statutory deadlines, and IT/GST notices. Includes AI agents for notice triage, GST reconciliation, filing reminders, and client communication. Comes with an AI Employee (Priya) who coordinates compliance work end-to-end.
Move candidates faster without cutting corners on fair hiring. Riley, your AI recruiting coordinator, screens each new candidate against the job's must-haves with must-have-by-must-have reasoning, coordinates interview scheduling by drafting availability requests, drafts honest candidate outreach and status updates, builds submittal packages for hiring managers or clients mapping experience to requirements, and gives the team a daily pipeline read. Screening never considers anything beyond skills and qualifications, and it never rejects a candidate or extends an offer; those stay human decisions. Everything else is draft-and-approve: candidate messages are prepared for a human to send, and interview times are proposed, never confirmed, by the agent. Four tables hold your job orders, candidates, interview scheduling, and pipeline metrics. A knowledge base holds your screening standards, submittal format, communication voice, and compliance guardrails, which every assessment and message follows. Automations included: new candidates are screened on arrival, plus an optional daily pipeline digest. Works out of the box with the roles and candidates you add; connect an ATS or job board later to sync candidates, and Gmail and Calendar to send outreach and confirm interviews from the platform. Best first step: replace the placeholder screening standards and compliance rules with your own and add an open job.
A complete outbound-to-pipeline sales engine for a small team. Define your ideal customer profile once in the ICP Profiles table and paste your product details into the Product Knowledge Base, then Sasha, your AI SDR, discovers and vets target accounts, enriches prospects, researches companies, scores leads, and drafts outreach, all grounded in your data. Inbound leads and your deal pipeline live in tables the founder can watch. Automations included: new prospects are auto-enriched on arrival, and an optional daily discovery run finds fresh accounts from your Active ICP. Requires an LLM provider plus the Apollo, Serper, Exa, and Google Search connectors. Replace the sample knowledge base content and the example ICP row with your own.
Answer more support tickets, faster, without losing the human touch. Sam, your AI support specialist, classifies every incoming ticket by category, sentiment, and priority, drafts an accurate reply grounded only in your knowledge base, and escalates anything it is not confident about instead of guessing. It matches recurring problems to approved canned responses, mines new tickets into known issues so answers stay consistent, tracks the feature requests buried in tickets, and produces a weekly insights digest of volume, deflection, and top drivers. Everything is draft-and-approve: replies are prepared for a human to review and send, and sensitive tickets (anger, churn, refunds, account deletion, legal or privacy) are always escalated. Four tables hold tickets, known issues, feature requests, and daily metrics. A knowledge base holds your help-center content and support voice, which every reply cites. Automations included: a reply is drafted the moment a ticket arrives, plus an optional daily sweep of anything still unanswered. Works out of the box with no external tool; connect Gmail later to send from the platform and Slack for the digest. Best first step: replace the placeholder knowledge base with your real help-center articles or URLs.
A complete, full-lifecycle sales engine for a growing team, spanning outbound prospecting and account management across one shared data layer, with a companion inbound concierge. Outbound: Sasha, your AI SDR, discovers and vets accounts, enriches prospects, scouts buying signals, researches companies, scores leads, and drafts cold email, LinkedIn, and nurture outreach. Account management: Maya, your AI Account Manager, qualifies opportunities, drafts and reviews proposals, flags pipeline risk, forecasts renewals and expansion, and keeps the CRM clean. Seven tables hold ICP profiles, target accounts, prospects, website leads, CRM contacts, opportunities, and customer accounts. Automations included: new prospects are auto-enriched on arrival, and an optional daily discovery run finds fresh accounts from your Active ICP. For the inbound engine, install the companion Website Sales Concierge team template, which routes visitor questions to product, pricing, and objection specialists and captures leads into the Website Leads table. Requires an LLM provider plus the Apollo, Serper, Exa, and Google Search connectors. Replace the sample knowledge base content and the example ICP row with your own. Outreach and proposals are drafted for a human to review and send.
