What to prepare, how to import it, how to improve recall quality, and how to maintain your materials over time

Knowledge Base Management

What this page solves

If you want answers to sound more like you and less like a generic assistant, this is one of the highest-leverage pages in the whole docs set.

Default recommendation

Our default recommendation is to build your resume in the cloud resume editor first, then sync that structured resume into the Knowledge Base. Clear project boundaries, responsibilities, outcomes, and timelines usually lead to better recall quality.

Most reliable import path right now

If you want the most reliable path, prioritize materials in this order: cloud resume JSON > project or system-design Markdown > Q&A docs in .txt or .md. The highest-leverage item is still the cloud resume.

Separate these two concepts first

The Knowledge Base itself is the retrieval layer that helps RAG use your resume, projects, and personal background. The notes panel is only a reading view for that material during the interview, so you can glance at it when you need support.

The Knowledge Base stores interview preparation materials so AI can reference your actual projects, resume facts, and preferred way of explaining things.

Remember these 3 rules first

Start with high-quality materials

A clean resume, project summaries, role-specific resume variants, and behavioral stories usually help more than dumping in random notes. A structured cloud resume is the default best starting point.

Structure matters more than quantity

Clear titles, sections, bullets, and project boundaries usually improve recall faster than simply uploading more documents.

Keep it aligned with the current role

The more your materials reflect the job you are preparing for right now, the more useful the outputs become.

Quick start

Open the Knowledge Base panel

Go to the Knowledge Base entry from the toolbar and confirm which materials are already present in your account.

Sync your cloud resume first, then add the first core set

Start with the structured resume from our cloud resume editor, then add your target-role resume variant, project summaries, and behavioral material.

Wait for parsing, then verify with a real prompt

After upload, preview the content and run one Ask or Voice flow to confirm the system is actually hitting the materials you expected.

Keep the reading panel open during the interview if needed

The Knowledge Base reading panel can stay open in stealth mode while you continue using live transcription Q&A and screenshot Q&A in the same round. This is especially useful for project follow-ups, resume facts, and behavioral examples.

Supported material types

JSON / structured resume

Best for project timelines, resume facts, responsibilities, and measurable outcomes because the structure is explicit. A structured resume exported from our cloud resume editor is usually the default best source.

Markdown

Great for project notes, system design outlines, behavioral stories, and technical review docs. Clear heading hierarchy usually leads to more stable indexing and recall.

TXT

Especially good for Q&A docs, prompt banks, and clean plain-text outlines. For a strict Q: / A: template, .txt and .md are both strong choices.

Word

Usually works well if the document uses clear headings and normal paragraph structure instead of heavy formatting tricks.

PDF

Usable, but parsing quality depends heavily on how the PDF was originally created and exported.

Overloaded mixed documents

If resume, project notes, prompt banks, and random meeting notes all live in one file, recall quality usually drops.

Unstructured scratch notes

Documents without titles, section breaks, or clear conclusions are much harder for AI to use reliably.

How to structure documents for better recall

Use clear project titles instead of vague labels like "project one" or "optimization work".
For each project, include context, responsibilities, technical approach, result, and measurable impact if possible.
Leave clear spacing between projects so the boundaries are easy to preserve.

The best default structure is Q: / A: with clear boundaries. Keep Q&A docs separate from resumes and project files instead of mixing everything into one source.
Group frequent follow-ups by theme, such as project follow-ups, behavioral follow-ups, or system-design follow-ups.
The current client generally works best with short questions, medium-length answers, and explicit keywords. In practice, questions around 6-20 words and answers around 20-200 words tend to work well.
If an answer is long, write the short main response first, then add 2-4 supporting bullets below it.

Suggested template:

Q: Give me a one-minute self introduction
A: I have 5 years of backend experience, and in the last two years I mainly worked on payment and transaction systems, focusing on high-concurrency optimization, stability work, and core API performance.

---

Q: Tell me about one representative project you led
A: I led a payment-query optimization project. I rebuilt the indexing and caching strategy, then separated slow-query paths from fallback paths, which reduced P95 latency from 420ms to 130ms and lowered peak CPU by about 35%.

Notes:
- Background: the payment-query path kept timing out under peak traffic
- Key actions: index redesign, hotspot caching, async handling for non-critical flows
- Result: lower latency, lower CPU, fewer alerts

---

Q: How would you answer “Why are you leaving your current role?”
A: I would focus on career direction and growth goals, explain that I want broader end-to-end ownership in a larger-scale environment, and still speak positively about what I learned in the current role.

User guidance:

Q&A docs are best for teaching the system how you usually explain things, while the cloud resume is best for grounding what you actually did.
If you only have time to improve one thing first, complete the cloud resume before writing more Q&A files.

Split by topic when possible: caching, queues, rate limiting, database design, and so on.
A "scenario - solution - trade-off - risk" structure usually produces the most speakable answers.
If you already have mindmaps or study notes, converting them into Markdown often improves results.

A "context - action - result - reflection" structure is usually the easiest to reuse in live interviews.
One project can support multiple story angles if the sections are kept separate.
Keep the real details: timing, partners, conflict, decision, and result.

Advanced settings worth watching

Smart chunking

Recommended by default. It tries to preserve project and paragraph boundaries so your materials do not get sliced too aggressively.

Context expansion

When one section matches, nearby content can be pulled in as support. This is especially useful for resume and project material.

topK

Controls how many segments are retrieved each time. Too low can miss details; too high can make answers wander.

How to maintain it before and after interviews

Before the interview

Keep only the materials that match the target role and current preparation track.

During the interview cycle

If answers feel generic, check structure and specificity first before uploading more and more files. The Knowledge Base reading panel can remain open as an extra support layer next to voice and screenshot workflows.

After the interview

Add follow-up answers, trade-offs, production incidents, and measurable outcomes that came up during the session.

Common questions

Why do JSON and Markdown usually work best?

Because titles, bullets, fields, and project boundaries are easier to preserve, so the system can tell which facts belong together.

Why do we recommend syncing the cloud resume into the Knowledge Base?

Because a structured resume keeps project boundaries, responsibilities, outcomes, and timelines much cleaner. Maintaining it in the cloud resume editor first, then syncing it into the Knowledge Base, usually improves recall quality more than uploading mixed documents manually.

What is the best way to write Q&A docs?

The best default is a standalone .txt or .md file using a Q: / A: template. Keep questions short, answers focused, and keywords explicit. Do not mix Q&A content into the same large file as resume and project material.

Why is PDF recall less stable?

PDF quality depends on original export settings and text-layer quality. If the file is important, converting it to Markdown or TXT often helps.

Why do answers still feel too generic?

Usually because the materials are too mixed, too vague, or too light on real details. Tighten the structure before adding more volume.