Most teams already have the answer somewhere. The problem is that it is buried across PDFs, handbooks, research papers, policies, and internal documents. Keyword search returns pages to read. A generic chatbot can answer without using the files that matter. Knowledge base question answering gives the AI a defined collection to search before it responds.
What is knowledge base question answering? It lets people ask natural-language questions across uploaded documents and receive answers grounded in those files, with source context available for verification and follow-up.
The key difference is grounding. Sharper answers from the knowledge base you select, shows the source context behind the answer, and surfaces related concepts so users can keep exploring instead of accepting a black-box response. You can build a reusable AI knowledge base for a trusted source collection, then ask questions across your knowledge base when you need a direct, verifiable answer.
In this walkthrough, we upload four files: two language-model research papers, the BASIS student handbook, and a sample employee handbook. Then we ask practical questions across the collection and verify the answers against the original documents.
How knowledge base question answering works
The workflow has seven observable stages:
- Ingest: Add the documents that should define the answer space.
- Parse: Turn the contents of each file into material the system can search while retaining its connection to the source.
- Index: Prepare that material for retrieval across the full knowledge base.
- Retrieve: Use the question to find the most relevant passages rather than asking a model to rely on general memory.
- Answer: Synthesize a direct response from the retrieved evidence.
- Cite: Connect claims to the supporting document and passage.
- Verify: Open the source context and decide whether the evidence supports the answer.
This sequence matters because a fluent response is not enough. A useful document answer should make its evidence inspectable.
Knowledge base Q&A vs. keyword search
Keyword search is useful when you know the phrase used in a document. It typically returns matching pages or snippets, leaving you to open results and assemble the answer. Knowledge base Q&A accepts a plain-language question, retrieves relevant passages even when the wording differs, and uses them to produce a direct response with source context.
The two approaches can complement each other. Search is a fast way to locate a known term; Q&A is better suited to questions that require explanation, comparison, or evidence from several passages.
Knowledge base Q&A vs. a generic AI chatbot
A generic AI chatbot may answer from broad training data or the immediate conversation. That is useful for open-ended ideation, but it does not guarantee that an answer reflects your current policy, handbook, or research collection.
Knowledge base Q&A narrows the task to the source set you selected. The resulting answer can be checked against those files, making the workflow more appropriate when provenance and current internal information matter.
The setup: four files, one knowledge base
The demo knowledge base contains four very different PDFs: BERT, RoBERTa, BASIS.pdf, and SampleEmployeeHandbook.pdf. That mixture matters because real teams rarely have one tidy source. They have research, policies, school handbooks, HR documents, and product files living side by side.
Step 1 - Upload files to build the knowledge base
The user uploads the four PDFs and imports them into Sharper. Sharper parses the files, indexes the content, and prepares the knowledge base for browsing and question answering.
Four uploaded files become one searchable knowledge base.
How Sharper works, in two moves. First, you build a knowledge base: upload your documents once and Sharper parses and indexes them into a reusable, searchable collection. Second, you ask: start a new task, select that knowledge base, and ask any question in plain English - as in the steps below. Sharper answers only from the selected knowledge base, so every response is grounded in your documents.
Step 2 - Browse documents, concepts, and graph views
After import, the user can browse the knowledge base in three complementary views. The document view shows each uploaded source. The concept view shows topics Sharper extracted from those files. The graph view maps how documents and concepts connect.
Document view: every source file is visible and ready for inspection.
Concept view: Sharper extracts topics from the uploaded knowledge base automatically.
Graph view: documents and concepts become an explorable map.
Step 3 - Click a concept to explore the topic
Concepts are not just tags. Each one has a summary, description, and source mentions. For example, the ParentSquare concept is summarized as the online portal for all school communications, payments, and the family directory.
Click a concept to see what Sharper knows and which source mentions it.
This makes the knowledge base useful even before a user asks a question. They can browse the topic map, inspect a concept, and jump back to the source that supports it.
Step 4 - Select the knowledge base and ask a question
Once the knowledge base is selected in the chat composer, the user can ask a direct question:
The question: "How many paid vacation days for the first year?"
Ask in plain English after selecting the knowledge base.
Sharper searches the selected knowledge base, finds the vacation policy in the employee handbook, and answers in context: full-time employees earn 2 weeks, or 10 days, of paid vacation during the remaining nine months of the first year, with no vacation earned during the first 90 days.
The answer is grounded in the source PDF, with the relevant passage visible beside it.
The answer also includes related concepts. Click one, such as Equal Employment Opportunity Policy, and Sharper opens a concept panel with a summary, description, and source mentions. That turns a single answer into a path for exploration.
Suggested concepts help users explore new topics without losing the answer context.
More Q&A examples from the same knowledge base
Because the knowledge base contains both research papers and handbooks, Sharper can answer very different questions while still grounding each response in the right source material.
Research question: Sharper explains how RoBERTa builds on BERT.
Policy question: Sharper answers with citations from the BASIS student handbook.
How to evaluate answer accuracy
Treat a cited answer as a reviewable result, not an automatic guarantee. Check five things:
- Source quality: Confirm that the knowledge base contains authoritative, current documents. An answer grounded in an outdated policy can still be wrong for today's decision.
- Missing evidence: Notice when the response or citation does not fully support a claim. The correct outcome may be “the available documents do not say.”
- Conflicting documents: If two sources disagree, inspect their dates, owners, and scope instead of accepting a blended answer.
- Retrieval coverage: Ask whether the cited passages cover the whole question, especially for comparisons or questions spanning several documents.
- Human review: Have a qualified person review answers used for legal, medical, financial, compliance, or other consequential decisions.
Accuracy improves when the source collection is maintained, questions are specific, and reviewers follow citations back to the original context.
Built for large document collections
The same workflow scales beyond four files. Sharper is designed for knowledge bases with thousands of large PDFs, with individual files up to 200 MB. Users can still browse the corpus, inspect concepts, explore the graph, and ask questions against the selected knowledge base.
That scale matters for teams working with employee handbooks, school policies, research libraries, legal archives, support documents, and operational manuals. The answer should not depend on someone remembering which PDF contains the rule.
What makes Sharper different
- Grounded in the provided knowledge base. Sharper answers from the files the user selected, not from generic model memory.
- Source context reduces hallucination. Answers are tied to retrieved passages, and users can inspect the original PDF context.
- Concepts make answers explorable. Related concepts help users move from one answer into the surrounding topic area.
- Multiple views support different workflows. Documents, concepts, and graph views let users browse before or after asking questions.
- Scale without losing precision. Sharper can handle thousands of large PDFs while still finding answers in context.
FAQ
What kinds of files can Sharper use for a knowledge base?
Sharper can work with large PDFs and common document formats, including research papers, student handbooks, employee handbooks, policy documents, manuals, and internal knowledge files.
How does Sharper reduce hallucination?
Sharper searches the selected knowledge base, grounds answers in retrieved passages, and shows source context so users can verify where an answer came from.
Can users explore topics instead of only asking questions?
Yes. Sharper builds document, concept, and graph views. Users can click a concept like ParentSquare or Equal Employment Opportunity Policy to see summaries, descriptions, and source mentions.
Does Sharper scale beyond a few demo files?
Yes. Sharper is designed to handle thousands of large PDF files, with uploads up to 200 MB per file, while still returning grounded answers from the selected knowledge base.
Can Sharper answer questions across unrelated files?
Yes. A single knowledge base can contain different document types, such as research papers and handbooks. Sharper searches the relevant source material for each question and answers in context.