--- title: "Knowledgebases: put your organisation's knowledge to work | Certant" description: "Point Certant at the documents and feeds you already have. It works out how they connect, and your agents and search run on the result, with the source page on every answer." url: "https://certant.ai/knowledgebases" --- The Certant MCP server is included on Business plans and above. Connect an agent you already use, such as Claude Code, and it can set Certant up for you. [See how it works](https://certant.ai/agents#mcp-server) # Put your organisation's knowledge to work Point Certant at the documents and feeds you already have. It works out how they connect, and your agents and search run on the result. [Book a demo on your documents](https://certant.ai/contact-sales) [Start free](https://certant.ai/register) More than 120,000 pages processed in production. ![The Certant knowledgebase builder](https://certant.ai/images/screenshots/kb-builder-canvas.png) Sources in Builds the model Embeds it Ready to ask The builder: sources in on the left, a searchable model out on the right. The problem ## Your wiki is hard to search Certant reads your pages instead of listing them. You get the answer with the page it came from, and it tells you when an older document says something different. [See what replacing a wiki looks like](https://certant.ai/replace-your-wiki) The same question, twiceExample "Do casual staff get long service leave?" A wiki search gives you a list - **Leave policy (OLD, do not use)**Edited 2019 - **Leave Policy FINAL v3**Edited 2021 - **HR handbook**Edited 2019 - **Meeting notes, 14 March**Edited 2023 And 212 more results. Certant gives you the answer Yes. Casual staff accrue long service leave after seven years of continuous service. The 2019 handbook says ten years and has been replaced. Enterprise agreement · clause 22.1 · page 63 Replaced: HR handbook 2019 What it makes possible ## Five jobs besides answering questions Most people first meet a knowledgebase through a chatbot. Once your documents are in, it can do a good deal more. Pick a job ### Turn a question into a table Ask for the liability cap in each of two hundred contracts and you get a table back, with a row for each contract and the page the figure came from. Liability cap, by contractExample | Contract | Cap, and where it says so | |---|---| | Vendor SOW 2026-114 | $2,000,000 Clause 9.2 · page 14 | | Facilities agreement | $500,000 Clause 12 · page 31 | | Software licence | Fees paid in 12 months Clause 8.1 · page 9 | | Cleaning contract | Not stated No cap found | If a contract doesn't state a cap, the row says that. ### It tells you where documents disagree Certant flags policies that contradict each other and shows you the two passages side by side. Conflict foundExample **Travel policy v3 · section 4.2** Claims must be lodged within 14 days of travel. **Expense procedure · section 2** Claims must be lodged within 30 days of travel. **Waiting** on the finance manager to say which one stands Asked nine times this month with no answer on file: parental leave for contractors. ### It watches what comes in Set a watch on anything in it, such as a contract coming up for renewal. When it trips, the person responsible is told and sent the clause. WatchExample **When** a contract is within 90 days of its renewal date **Checked** whenever new data lands **Tell** the contract owner, with the renewal clause attached Facilities agreement · clause 3.1 · page 6 ### First drafts from your own record Certant drafts tender responses, grant applications, submissions and board papers from what your organisation has already said and decided. Tender response · 4.2 Safety recordExample We have completed eleven projects of this kind since 2021 without a lost-time injury. Our safety management system was last audited in March 2026. Project register · page 4 Audit report, March 2026 · page 2 ### It remembers how things were decided Decisions and exceptions are kept with the reasons for them, so you can ask how something was handled last time. PrecedentExample "Has a member ever had fees waived after a natural disaster?" Yes, twice. The board approved a twelve-month waiver in January 2020 and again in March 2022, both times for members in a declared disaster area. Board minutes, 14 March 2022 · item 6 Board minutes, 22 January 2020 · item 3 The same knowledgebase sits behind [chatbots](https://certant.ai/chatbots), the [Agent Builder](https://certant.ai/agents), [Analytics](https://certant.ai/analytics), [Strata](https://certant.ai/strata) and the [AI FDE](https://certant.ai/ai-fde). When you correct the knowledgebase, each of them uses the correction. Agents ## Agents you can trust to act Most Certant customers use their knowledgebase behind an agent, to back up each decision it makes. ### What the knowledgebase supplies **Knows who is who.** The supplier, the contract party and the member record are treated as one organisation. **Calculates figures.** Numbers are worked out from the table they sit in. **Follows your rules.** Delegations and limits come from your own policies. **Uses the right version.** It applies the rule in force on the date in question. **Remembers decisions.** Past decisions and corrections stay in the knowledgebase. **Stops when it doesn't know.** It hands the job to a person and says what was missing. ### What the agent does with it Example 1. **Harlow Civil Pty Ltd.** The same supplier as "Harlow Civil" in the accounts and the party to contract 2026-114. 2. **Invoiced $18,400** against a contract rate of $17,900. Over by $500. Contract 2026-114 · schedule B · page 22 3. **Needs the finance manager.** Payments over $10,000 are not the agent's to approve. Delegations policy · section 3.2 · page 5 4. **Rate as at 12 May 2026,** the invoice date. The schedule changed on 1 July and was not used. 5. **Approved once before.** A similar variance was accepted in February, with a written reason. Approval record · 9 February 2026 6. **No purchase order on file.** Stopped, and sent to a person. Waiting on approval Agent run: invoice approval. Each step uses the numbered item beside it. [See the Agent Builder](https://certant.ai/agents) [How agents share one knowledgebase](https://certant.ai/multi-agent-workflows) Who uses it ## What people build with it Research ### A research corpus you can ask A water authority asks its whole research library one question and gets one cited answer. Compliance ### Two rule sets, side by side A bank loads a new prudential standard beside its policies and sees every clause that changed. HR ### HR answers without the queue An association's employees answer their own leave and benefits questions. **Ask Certant to build it.** Describe the knowledgebase you want and the AI FDE builds it, then waits for your approval. [See the AI FDE](https://certant.ai/ai-fde) The detail ## Read how it works Open a heading to read that part. The longer version ## What it does, in full ### Why a wiki is hard to search Ask anyone who has used a wiki for a few years. The answer is in there somewhere, under a title someone picked in 2019, next to three older versions that disagree with it. Certant reads those pages instead of listing them. You ask the question and get the answer with the page it came from, and it tells you when an older document says something different. It reads SharePoint, Confluence and your shared drives together, so it doesn't matter where a document was filed. ### Turn a question into a table Ask for the liability cap in each of two hundred contracts and you get a table back, with a row for each contract and the page the figure came from. Search works on meaning, so it finds the clause whatever the author called it. Every member, contract, site and policy also gets a page of its own, written from everything that mentions it. ### It tells you where documents disagree Certant flags policies that contradict each other and shows you the two passages side by side. It does the same for procedures that have been replaced and guidance that was saved more than once. It also keeps a list of the questions it couldn't answer, which shows you what your documents are missing. ### It watches what comes in You don't have to upload everything by hand. Certant takes file drops, scheduled exports, data sent through the API and forwarded email, and it checks each delivery and keeps it on file. You can set a watch on anything in it, such as a contract coming up for renewal or a new document that contradicts a policy. When it trips, the person responsible is told and sent the clause. ### First drafts from your own record Certant drafts tender responses, grant applications, submissions and board papers from what your organisation has already said and decided. Each paragraph shows the source it was drawn from, so whoever reviews it can check each claim against the original. ### It remembers how things were decided Decisions and exceptions are kept with the reasons for them and the documents they relied on, so you can ask how something was handled last time. That still works after the person who made the decision has retired. ### What an agent gets from the knowledgebase Access to your systems lets an agent act. What it reads in your knowledgebase is how it decides whether to, and the example below shows the page behind each step. **Knows who is who.** The supplier in your accounts, the party to the contract and the member record are the same organisation, and the agent treats them as one. An agent without that matches on the name. **Calculates figures.** Numbers are read from the table they sit in and worked out from there. The AI doesn't write them. **Follows your rules.** Delegations, limits and escalation paths come from your own policies. An agent without them only knows what somebody thought to put in its instructions. **Uses the right version.** It applies the rule that was in force on the date in question and tells you which one that was. **Remembers decisions.** Past decisions and corrections stay in the knowledgebase, so you don't have to make the same correction again for the next agent. **Stops when it doesn't know.** If the knowledge isn't there, the agent hands the job to a person and says what was missing. ### What people build with it **A research corpus you can ask.** A water authority indexes its whole library of research. "What do we know about PFAS in NSW?" comes back as one answer across many papers, each with its citation. **Two rule sets, side by side.** A bank loads the new prudential standard beside its existing policies. The model surfaces every clause that changed and every gap, with the page each came from. **HR answers without the queue.** An association loads every benefits, leave and procedure document. Employees answer their own questions, with a link to the exact policy section. ### Ask Certant to build it The AI FDE sits in a drawer on every screen. Describe the knowledgebase you want in plain English and it builds it, then waits for your approval before anything goes live. How it is built ## How a file becomes an answer The same six steps the product runs, in the order a document travels through them. **Bring in your documents.** Drag files in, or pull from S3, SharePoint or Confluence. PDFs, Word, slides, spreadsheets, scans and images. **Read every file properly.** A different reader for each kind of file, so tables stay tables and handwriting still gets read. The four readers are listed further down. **Split it sensibly.** A table or a clause is never cut in half. **Match on meaning.** Two passages that say the same thing in different words are found as related, not just ones that share a word. You can bring your own model for this. **Work out how it connects.** People, employers, clauses, sites, claim types: named and linked to each other. Let Certant decide the types, define them yourself, or enforce them strictly. **Reconcile duplicates.** "Staff", "Worker" and "Employee" end up as one thing rather than three. What comes out ## What Certant builds from your documents Certant reads every passage and pulls out the things that matter: people, organisations, clauses, salary bands, claim types. It then reconciles them, so "Staff", "Worker" and "Employee" become one searchable thing. **You choose how strict it is.** Let Certant decide the types, define them yourself, or enforce them. **Duplicates get reconciled.** One entity at a time, with a lock, so two documents describing the same thing end up as one. **The rest of Certant uses it.** Search, chatbots, agents and dashboards read from this model. Knowledge graph ![A knowledge graph built from uploaded documents](https://certant.ai/images/knowledgegraph-preview.png) Entity browser ![Browsing one entity and its relationships](https://certant.ai/images/graph-preview.png) What comes with it ## What comes with a knowledgebase Reranking Weak matches are filtered out before they reach the model. The cost is attributed per knowledgebase, and the providers are listed under Specifications below. PageIndex Certant takes a quick first pass to narrow the field before retrieval starts, so a large knowledgebase doesn't slow a question down. PDF visualiser Click a citation to open the exact page with the passage highlighted. Relevance can be scored sentence by sentence. Storage backends File-based storage for a single user, or PostgreSQL for production. Both are listed under Specifications below. ![A citation opened at the page, with the passage highlighted](https://certant.ai/images/pdf-visualiser-preview.png) Click a citation, land on the page. ![Choosing how results are ranked](https://certant.ai/images/reranking-preview.png) Ranking is configurable per knowledgebase. For the CIO ## How it is governed Access follows the person People, chatbots and agents see only the knowledgebases and documents they are entitled to. [Security](https://certant.ai/security) Sources shown Answers, figures and agent steps show the document and page they came from. [Architecture](https://certant.ai/architecture) Obligations register Obligations in contracts, standards and agreements are pulled into one register and matched to the policies that meet them. [Security](https://certant.ai/security) Personal information See where personal information sits across your documents, and find everything held about one person when a request arrives. [Security](https://certant.ai/security) Corrections on the record A correction is kept in a ledger and replayed whenever new documents arrive. Undo is a new entry, never a deletion. [Architecture](https://certant.ai/architecture) Audit and retention Certant records who asked, what was read and what an agent did, and shows you how long each record is kept. [Security](https://certant.ai/security) For your IT team ## What your engineers will ask about These are the components by name. You can change any of them in settings. Certant picks a reader for each kind of file, and you can change which one it uses. | Reader | Best on | What it is | |---|---|---| | MinerU | PDFs, scans and complex layouts | GPU-grade layout and table extraction, local or hosted, with pipeline, VLM and hybrid backends. | | Docling | Word, slides and structured PDFs | IBM's structured document parser, on a local CPU or a remote API, with optional Granite Vision for picture descriptions. | | PaddleOCR | CJK, right-to-left and mixed-script documents | Multilingual OCR with chart and formula recognition, and a configurable layout threshold. | | GLM-OCR | Offline and sovereign deployments | MIT-licensed open-source vision OCR that runs offline, with an optional faster variant. | You can set the mode for a query, a chatbot or an agent. | Mode | Use it for | How it works | |---|---|---| | Naive | Simple factual questions | Vector lookup only, which is the fastest path. | | Local | Precise follow-ups that depend on context | Walks the knowledge graph one hop from named entities. | | Global | Strategic, big-picture questions | Aggregates across the entire graph. | | Hybrid | Questions that need depth and breadth | Local and global, combined in one pass. | | Mix Default | Most questions | Graph context fused with vector chunks, then reranked. | | Mix-concise | Short operational replies | The same evidence as Mix, with a brief answer. | | Conversational | Live chat | Keyword search (BM25) fused with vector search by reciprocal rank fusion, with no entity extraction. | | Bypass | Questions about the knowledgebase itself | The model alone, with no retrieval. | | Setting | Detail | |---|---| | Chunk size | Token windows from 128 to 4,096, or structure-aware splitting. | | Embedding model | Qwen3-Embedding-4B by default, 2,560 dimensions, batched. You can bring your own. | | Graph merge | A single sequential worker with distributed locking, and model-assisted summarisation when an entity's descriptions pile up. | | Storage | File-based (NanoVectorDB and NetworkX) for one user, or PostgreSQL with pgvector and Apache AGE for production. | | Reranking | Cohere, Jina, Aliyun or DeepInfra, with cost attributed per knowledgebase. | | Your own agent | An MCP server gives an agent you already run the same tools, behind the same permissions and approval gates. | The evidence ## What independent research found Four published studies support this, and a fifth shows where a knowledgebase makes no difference. Each one is linked. Agents struggle with written rules In a published test of AI agents working under written rules, the leading models failed more often than they succeeded, and gave different results on repeat runs. [Yao and others, 2024](https://arxiv.org/abs/2406.12045) Wrong figures are hard to spot dbt Labs found that when an AI works out a figure unaided and fails, it returns a plausible wrong number. When it asks through a checked model of the data and fails, it returns an error. [dbt Labs, 2026](https://docs.getdbt.com/blog/semantic-layer-vs-text-to-sql-2026) Accuracy falls with more text Chroma found that accuracy fell in every AI model it tested as the model was given more to read, even on simple tasks. [Chroma, 2025](https://www.trychroma.com/research/context-rot) Precedent speeds up resolution LinkedIn connected its past support tickets into one model and reported faster resolution over six months in its customer service team. [Xu and others, 2024](https://arxiv.org/abs/2404.17723) Where it makes no difference Han and others found that plain search does better on some kinds of question and connected knowledge on others. Certant chooses between them for each question and answers from the source text either way. [Han and others, 2025](https://arxiv.org/abs/2502.11371) Try it on your documents These studies used other people's documents. Bring a folder of your own and a job you do by hand, and see how it does on those. [Book a demo on your documents](https://certant.ai/contact-sales) ## Your documents can stay in the building Every reader, model and store on this page has a version that runs inside your own walls. [Book a demo on your documents](https://certant.ai/contact-sales) [Start free](https://certant.ai/register) Certant Cloud, your cloud, or fully air-gapped. AnswerExample Long service leave accrues after seven years of continuous service, and can be taken in two blocks with the employer's agreement. Enterprise agreement · clause 22.1 · page 63 Leave policy v4.pdf Every answer opens the document at the page it came from.