Knowledge Base Assistant

Every organization has knowledge scattered across documents, wikis, shared drives, and people’s heads. When someone needs an answer, they search through folders, ping colleagues on Slack, or give up and figure it out themselves. A knowledge base assistant makes all of that information accessible through conversation.

The Problem

Information exists but it’s not accessible. Your team has documentation, process guides, policy documents, strategy decks, training materials, and years of institutional knowledge. But finding the right information at the right time is a full-time job nobody signed up for.

The result: people reinvent solutions that already exist, make decisions without context they could have had, and spend more time searching for information than using it.

Traditional knowledge bases help, but they require the user to know what to search for. If you don’t know the exact term, the right document title, or which section contains the answer, search fails.

How the Softcery Platform Solves It

Build a knowledge base assistant that thinks, connects dots, and explains across your entire knowledge base.

More Than Search

The Knowledge Base Assistant preset creates an agent that approaches questions like a knowledgeable colleague. It doesn’t just find the closest document match – it considers what’s directly relevant, what’s tangentially useful, and what connections might not be obvious.

Ask “What’s our refund policy?” and it gives you the policy. Ask “What should I consider when a client requests a refund after 60 days?” and it draws from the refund policy, the exception handling process, and client relationship guidelines – synthesizing an answer that no simple search would return.

Deep Knowledge, Smart Chunking

Upload everything your team needs access to:

• Policy documents – HR policies, operational procedures, compliance guides
• Product documentation – Technical docs, API references, user manuals
• Strategy documents – Business plans, market research, competitive analysis
• Process guides – Onboarding workflows, approval processes, escalation procedures
• Training materials – Courses, workshops, best practices

The platform’s markdown-aware chunking preserves document structure. A policy document with sections on different leave types is chunked so each section remains coherent – the agent retrieves “Parental Leave Policy” as a unit, not a fragment that starts mid-sentence.

Transparent Sourcing

Every answer shows where the information came from. The inspection panel reveals which documents were retrieved, their relevance scores, and the exact content the agent used. This matters for knowledge bases where accuracy and source authority are important – you can always trace an answer back to the original document.

What It Looks Like in Practice

Team member: “What’s the process for approving vendor contracts over $10k?”

The agent retrieves the procurement process document, the approval chain for that spend level, and any relevant contract templates or checklists. It explains the process in conversational terms – step by step, with the specific thresholds and approvers for their situation.

New hire: “How does our deployment pipeline work?”

Instead of pointing to a 30-page architecture doc and saying “it’s in there somewhere,” the agent explains the pipeline – from code commit through CI/CD to production – pulling from technical documentation and supplementing with context about why the pipeline is designed the way it is.

Manager: “What’s our policy on remote work for international employees?”

The agent draws from HR policies, international employment guidelines, and any country-specific documentation, synthesizing an answer that covers the relevant policies while noting areas that might need HR consultation.

Configuration Breakdown

Component	Setup
Behavior	Knowledge Base Assistant preset with your organization’s context
Knowledge	Policy docs (file upload), process guides (text), internal wiki (website crawl), training materials (file upload)
Evaluations	Factual accuracy (block), scope enforcement (warn)
Channel	Password-protected link for internal teams, or embed in your internal tools
Model	Claude Sonnet 4 for thoughtful, synthesized answers
Advanced	Temperature 0.4 (accuracy over creativity), retrieval limit 30 (broad search), similarity threshold 0.15 (catch related content)

Who This Is For

• Growing companies where institutional knowledge is outpacing what new hires can absorb
• Distributed teams where “ask someone who knows” doesn’t work across time zones
• Regulated industries where finding the right policy quickly matters for compliance
• Any team where the answer exists in a document but nobody can find it fast enough

Internal vs. External

The same approach works for both internal and external knowledge bases:

Internal: Password-protected link channel, loaded with company policies, processes, and institutional knowledge. Accessible to your team only.

External: Embed widget on your product’s help center, loaded with product documentation and user guides. Accessible to your customers.

The configuration is identical. The difference is the knowledge you load and the channel access you configure.

Building Institutional Memory

Over time, the knowledge base assistant reveals gaps in your documentation. When the agent can’t answer a question – you’ll see it in the conversations – that’s a signal that the information either doesn’t exist in documented form or isn’t well-structured enough to be retrieved.

Each gap is an opportunity: write a text source with the answer. Over time, your knowledge base grows from what your team actually asks about, not from what someone guessed would be useful to document.

Getting Started

1. Create an agent and choose the Knowledge Base Assistant preset
2. Upload your most-referenced documents as file sources
3. Add your internal wiki as a knowledge source if it’s web-accessible
4. Set up factual accuracy evaluation to prevent hallucinations
5. Create a password-protected link channel for your team
6. Share the link and password – start collecting real usage data