Your embedding model
Any model deployed for embeddings. A question is always embedded with the model its library was indexed with.
Agentic RAG
Agentic RAG: the agent decides where to look.
Classic RAG searches once and hopes. Agentic RAG chooses how to search, checks what came back, and cites what it used.
Botminds agents answer from your own libraries, documents, records and the web. Built for a project and connected to knowledge bases of millions of documents, each agent picks from more than two dozen ways to search, reads for each person only what that person may read, and shows the page behind every answer.
Question“What notice period applies to contractors in Spain?”
The agent decidessearch or notwherewhich way to searchagain, if it missed
Knowledge librariesmillions of documents, many at once
Project documentsall of them, or one saved view
Recordsfiltered and counted, not guessed
Pages and summariesread word for word, or summarised
This conversationthe files you dropped in
The public websearched and cited by address
Your read rulesapplied to every search made for you — a collection you may not read never comes back
PassagesS1S2S3S4
Answercites S1 and S3 — only what it used
What is agentic RAG?
Agentic RAG is retrieval augmented generation in which an AI agent, not a fixed pipeline, runs the retrieval. The agent decides whether a question needs sources, which sources to search, which way to search them, and whether the passages it found actually answer the question — searching again when they do not. Classic RAG retrieves once for every question and hands the model whatever came back.
Scales out on KubernetesServices and workers add capacity as work arrives, and give it back when it is done.
Built for millionsMillions of documents, records and agent runs — the indexes and workers grow with them.
Any cloud, any regionAzure, AWS, Google Cloud or on-premises, in the region your data must stay in.
Always onRedundant services, rolling releases and continuous backups, for mission-critical work.
Secure and governedSingle sign-on, encryption, least privilege and a sealed audit trail.
Agentic RAG vs RAG
Agentic RAG vs RAG: who decides when to search.
Retrieval augmented generation gives a language model passages from your documents so it answers from them instead of from memory. The classic design wires that into a fixed pipeline. The agentic design hands the retrieval to the agent.
Classic RAG · a fixed pipeline
- Questionevery question is searched, needed or not
- Retrieve oncethe nearest passages from one index
- Stuff the promptwhatever came back, relevant or not
- Answerfrom the passages, or from memory when they miss
One shot. A miss is answered anyway.
Agentic RAG · a loop the agent drives
- Plandoes this need our sources at all?
- Pick the source and the searchwhich library, which documents, by meaning or exact words
- Read and checkdo these passages answer it?no → search again, narrower
- Answer and citeonly the passages it used — or say the sources do not cover it
↺ as many searches as the question needs
Not two ways to search — many
Every way an agentic RAG agent can search.
Enterprise knowledge is libraries, live documents, tables of records, the file someone just dropped in and the public web. One index searched one way cannot serve all of it. A Botminds agent carries a toolkit of searches and picks the right one for each step — often several for one answer.
28
ways to search, one agent.It picks the way per step — and combines them in one answer.
01Find the passage
- By meaningthe passage that says it in other words
- By exact wordsIDs, clause numbers, quoted phrases
- Many libraries at onceone search over every library the agent holds
- Project-wideevery document in the project at once
- Within a saved viewthis year’s claims, one client’s files — the view the agent is bound to
- By labelregion = EU and product = Motor — any mix of your own labels
- Named documents, pages, a sheetthe contracts it names, pages 12–14, one tab of a workbook
- Grouped by documentside by side per document — to compare, or find where they conflict
02Find the document
- By a half-remembered titleresolves the name a person half recalls
- Which ones mention itranked by how many passages match
- A whole document setevery member listed, the ones that mention it marked
- The newestmost recently added first
03Read it
- Pages, in orderthe exact pages, word for word
- Page summariesa short summary of every page, indexed beside its text
- Document summarywhat a document is, not only what it says
- Page mapthe outline of a long document, to choose where to read
- Scanned pagesa vision model reads pages that have no text layer
- Across many documentsread and summarised in parallel, then combined into one answer
04Query the records
- Typed filters and sortingprecise subsets of a table, on real columns
- Counts and statisticstotals, medians and percentiles per group, worked out in the database
- Questions as SQLread-only queries, checked before they run, with charts and spreadsheet export
- Records in your own wordsthe column people read is searched; the rest become filters
- Documents by their fields“claims over $10,000 from March”, filtered on extracted values
- Against its peersone value beside its group’s median, small groups withheld
05Beyond the library
- Files dropped in the chatread and searched inside that conversation
- Search the webcurrent facts, cited by address
- Read a web pagethe page at a given address, fetched and read
- Your own systemsan API or MCP server the agent is granted
Every way, the same wallsyour workspace›the agent’s libraries and view›what the asker may read›cited to document and page
Three ways, one answer
You
Which supplier contracts mention clause 14.3, and how much did we spend with those suppliers this year?
Agent
Searching by the exact reference “14.3” finds it in three contracts — Northwind Logistics, Bayside Freight and Arden Packaging — where it is the termination-for-convenience clause. Counting the invoice records for those three suppliers: $1.42 million this year, $860,000 of it with Northwind. Sources: three contracts, p. 9, p. 11, p. 7 · Invoices table
- Exact wordsfound the clause number that a search by meaning would blur.
- Find the documentslisted which contracts carry it.
- Counts in the databasesummed the spend, instead of estimating it from passages.
Enterprise RAG at scale
Knowledge bases of millions of documents.
A knowledge base on Botminds has no fixed ceiling. Indexing workers scale out on the queue, each search index is created with its own shard and replica settings, a project can keep its knowledge in an index of its own, and agents built for one project connect to libraries of millions of files — on Azure, AWS, Google Cloud or on-premises.
01 · ArriveDocuments pour inuploads, connectors, mailboxes, channels
- bursts land on a durable queue
- nothing is dropped while workers catch up
02 · IndexWorkers scale outmore workers as the queue grows, fewer as it drains
- documents indexed in parallel
- embedding calls batched and run at once
03 · StoreIndexes set per projecteach search index created with its own shard and replica settings
- a project can keep its knowledge in an index of its own
- shared or dedicated, chosen in the knowledge settings
04 · ServeAgents searchthousands of agents and conversations at the same time
- a library is one tag on each passage, never a list of its documents
- answers grounded in the passages they cite
Re-index without downtime
Document by document, newest pass wins — the rest of the library keeps answering, and agents are told how much is ready.
Millions of documents
No fixed ceiling on a knowledge base. Agents built for a project connect to libraries of millions of files.
Any cloud, any region
Azure, AWS, Google Cloud or on-premises — the index stays where your data must stay.
From document to knowledge
Chunk, embed, index — and count before calling it ready.
Every document added to a knowledge collection is cut into passages, embedded and indexed automatically, with no pipeline to wire. The step most systems skip is the last one: proving the index holds what the run says it wrote.
- Addupload, a connector, an email — any way in
- Extracttext per page; switch on a vision model to read scanned pages
- Chunkpassages sized for the text; table pieces repeat their header row
- Embedin batches, with the collection’s own model
- Indexeach passage stamped with its collection and pass
- Verifycount what the index holds; it must match
Every document ends in one of three states
Readythe count matched — pipeline success alone never sets it
Processingstill running, or the index has not caught up yet
Failedwith the cause in words — “no extractable text”
Summaries to find by
Switch on page and document summaries, indexed beside the text, for questions about what a document is.
Scanned pages read
Switch on a vision model and pages with no text layer are read like any other.
Your own labels
Stamp a service, a region or a product on every passage, so a search can be narrowed to any mix of them.
Keeping knowledge current
Re-indexing without stale answers.
Policies change, files are replaced, retries overlap. The danger in a knowledge base is an old version still being quoted. Every pass over a document is numbered, and only the newest one counts — while the rest of the library keeps answering.
One document, three passes
Pass 1The first version of the policy, indexed in 18 passages.purged by pass 3
Pass 2A retry that stalled and finished late, after pass 3 began.cannot overwrite pass 3
Pass 3The edited policy, 16 passages, counted in the index.the only version served
Document deletedits passages leave the index, with its status — a deleted policy is never quoted
Settings changeda full re-index of the collection, run once someone confirms the cost
Scheduled checkcollection, status and index are compared on a schedule you set, and any drift is repaired
The agentic part
How an agentic RAG agent decides when and where to search.
The agent is told what each of its sources holds and when to use it. On every question it makes four choices, and each one can branch.
01Does this need our sources?
- a greeting or a follow-up: no
- a fact, a rule, a figure: yes
02Where to look?
- knowledge libraries
- project documents, or a view
- records
- files in this chat
- the public web
03How to search?
- by meaning, or by exact words
- narrowed by label, document, page or sheet
- find the documents first, then read them
- count it in the records
04Do the passages answer it?
- yes → answer and cite
- partly → search again, narrower
nosay the sources do not cover it
Coverage check
Switch on the knowledge tools and a search whose passages probably do not answer the question comes back with a note, so the agent searches again or says so.
Knowledge lock
Choose to have the agent answer from its sources alone, and say plainly when they do not cover a question.
Failure said as failure
A search that fails says it failed, and a library still indexing says how much of it is ready — never an empty result that reads as “nothing there”.
Permission-aware retrieval
Permission-aware retrieval: the asker’s rules, inside the search.
An agent that can read more than the person asking is a leak with a chat box. On Botminds the walls are applied inside every search, in every one of its ways — so nothing outside them ever reaches the model.
Your workspacethe tenancy wall — no search ever crosses it
This project, plus the libraries the agent was givena knowledge library it was not given stays out, even in the same project
The agent’s viewwhen it is bound to one, only the documents that view selects
What the person asking may readwhen a person is asking, their read rules — a collection refused to them is refused to their agent
The passages that come backinside every wall at once
Plus: a file attached to a conversation stays inside that conversation. A run with no person behind it — an API key, email or Teams — searches inside the third wall, so you choose what a reachable agent holds. And when the rules cannot be checked, the search uses the last confirmed answer briefly, then fails out loud — a failure is never dressed as “nothing found”.
One rule, screens and search
A collection’s read rules — which roles may read it, and which rows — apply to the screens, to an agent’s search and to its queries over records, whenever a person is asking.
Libraries are given
A knowledge library is searched by the agents it is attached to — in its own project or another project of the same workspace.
You choose what a reachable agent holds
An agent answering by email, in Teams or through an API key searches what it was given. Attach what everyone who can reach it may read.
Answers you can check
Citations: the sources the answer used.
Listing every passage a search returned looks thorough and misleads. Botminds agents mark the passages they relied on, and those are the sources shown — each with its document and page.
- Labelled
Each passage reaches the agent with a label: S1, S2, S3.
- Marked
The agent names the labels it relied on.
- Kept
Those passages become the sources; with none marked, the closest matches.
- Shown
Clean prose, and sources that open beside the chat at the page.
A grounded answer, as the reader sees it
You
What notice period applies to contractors in Spain?
Agent
Thirty days, unless the statement of work sets a longer period. The Spain addendum overrides the global 14-day default. Sources: Contractor Policy — Spain addendum, p. 3 · Global Contractor Policy, p. 11
You
And in Portugal?
Agent
The sources do not cover Portugal. The global policy’s 14-day default may apply, but no document confirms it for Portugal.
Side by side
Classic RAG vs a vector database vs Botminds agentic RAG.
Teams building enterprise RAG usually start from a vector database and a retrieval script. The gap between that and a system people rely on is the rows below.
| Classic RAG pipeline | Vector database + your code | Botminds agentic RAG | |
|---|---|---|---|
| Who decides to search | The pipeline, every time | Your code | The agent, per question |
| Ways to search | Usually meaning | Whatever you build | More than two dozen: meaning, exact words, labels, documents, pages, summaries, records, chat files, the web |
| Scale | One index you size | Yours to shard and run | Workers that scale out, indexes set per project, millions of documents |
| Ready means | The job finished | Whatever you check | The index count matches what the run wrote |
| Re-indexing | Rebuild, or risk old passages | Your code | Numbered passes; newest wins; library stays online |
| Who may read what | One shared index | Filters you must remember | Workspace, libraries, view and the asker’s read rules on every search |
| Citations | Every passage retrieved | Your code | The passages the answer used, to the page |
How to judge any RAG platform. Ask four questions: when is a document actually searchable, what happens to the old version after an edit, whose permissions does the search run under, and what does the answer say when nothing relevant was found.
Questions people ask
Agentic RAG, answered.
More from the platform: the full architecture, and how security works at Botminds.
What is agentic RAG in simple words?
Agentic RAG is retrieval augmented generation where an AI agent runs the search instead of a fixed pipeline. It decides whether to search, where and which way, checks whether what it found answers the question, and searches again or says so when it does not.
What is the difference between agentic RAG and traditional RAG?
Traditional RAG retrieves once for every question, from one index, and passes whatever came back to the model. Agentic RAG lets the agent choose the source and the way to search, run several searches, and judge the results before answering.
How many documents can a Botminds knowledge base hold?
There is no fixed ceiling. Indexing workers scale out on the queue, each search index is created with its own shard and replica settings, a project can keep its knowledge in an index of its own, and a library is picked out by one tag on each passage rather than a list of its documents — so agents built for a project connect to knowledge bases of millions of documents.
What does an enterprise RAG architecture need beyond a vector database?
Many ways to search, a verified indexing pipeline, clean replacement of old versions, permission checks inside every search, an index that scales out, citations, and honest failure. The vector database is one part of that.
How does permission-aware retrieval work on Botminds?
Every search is held to the workspace, to the project and the libraries the agent was given, to the agent’s view, and — when a person is asking — to what that person may read under each collection’s read rules. See AI governance.
Does Botminds use hybrid search?
Agents search by meaning and by exact words under the same filters, and choose per search — or run both for one question. They can also narrow by label, document, page or sheet, and count in the records instead of searching text.
Can I chat with documents and see where the answer came from?
Yes. Answers show the sources they used — document and page — and a source opens beside the chat. The passages the answer marks as used are the ones listed.
How do I build a knowledge base AI agent on Botminds?
Create a knowledge collection, add documents by upload or a connector, and attach it to an agent built for your project. Piper, the platform assistant, can do each step on request. See enterprise vibe coding.
What happens when a document is updated or deleted?
An update starts a new numbered pass; older passes are removed and a late result from an overtaken pass cannot overwrite the new one. A deleted document leaves the index with it. The rest of the library keeps answering throughout.
Is agentic RAG the same as context engineering?
No. Retrieval finds material in your sources. Context engineering decides everything else the model sees on a turn — who is asking, the page, the conversation, the instructions. See context engineering.
Learn the basics
Agentic RAG vs RAGHow retrieval changes when an agent decides when, where and how to search→Keyword vs semantic searchBM25, embeddings, hybrid search and re-ranking, and where each one fails→Knowledge graph vs vector databaseRelationships against similarity: what each stores, answers and costs→RAG vs fine-tuningWhich fixes missing knowledge, which fixes behaviour, and when to use both→All guidesEvery explainer, by topic→
Answers from all your knowledge, with sources.
Bring the library your team searches every day. We’ll index it, attach it to an agent and show you every answer’s sources, under your own permissions.