Knowledge base recall quality directly determines the answer quality of AI interview assistants. When your project experience is fragmented into pieces, AI cannot generate convincing answers.

We discovered that Markdown resumes had only 0% project completeness rate. This means when the interviewer asks "Tell me about your project experience at ByteDance," AI can only recall the project name without key information like background, solution, and results.

This article shares how we improved Markdown resume project completeness from 0% to 93%, reaching industry-leading levels.

Why Recall Quality Matters So Much

In RAG (Retrieval-Augmented Generation) applications, recall quality is the decisive factor. Research from Google, OpenAI, and other tech giants shows that for every 10% improvement in recall quality, final answer quality improves by about 15%.

When analyzing user feedback, we discovered a critical issue: when AI answers project experience questions, it often "knows the project name but not the details." This leads to unconvincing answers that fail to showcase the candidate's true capabilities.

After deep diagnosis, we found the root cause lies in chunking strategy: project experience is fragmented into pieces, preventing AI from accessing complete project context.

Industry Pain Points: Why Existing Solutions Fall Short

We surveyed knowledge base products on the market and found they universally suffer from these issues:

Problem 1: Simple Character-Based Chunking

Most products use fixed character count chunking (e.g., 500 chars, 800 chars), completely ignoring semantic boundaries. This leads to project titles separated from content, key information fragmented, and missing context during recall.

Problem 2: Lack of Format Adaptation

Different document formats (JSON, Markdown, PDF) have vastly different structures, but existing solutions use the same logic for all formats, resulting in JSON resume field relationships destroyed, Markdown headers chunked separately, and unstable PDF parsing quality.

Problem 3: Missing Semantic Understanding

No recognition of key entities in documents (projects, QA pairs, tech stacks), unable to guarantee completeness of these entities.

We decided to fundamentally refactor the chunking strategy, drawing from Google Search's semantic chunking and best practices from major tech companies' RAG systems.

Our Solution: Intelligent Semantic Chunking

Core Principle: Semantic Completeness First

We proposed the "Semantic Completeness First" principle, aligned with Google's "understanding context" philosophy in BERT. The core idea: maintaining semantic unit integrity is more important than pursuing chunk size.

Technical Implementation: Three Major Innovations

Innovation 1: Multi-Format Adaptive Recognition

We developed an intelligent document type detection system that automatically recognizes resume documents (JSON, Markdown, TXT, Word, PDF), QA documents (question-answer format), and technical documents (architecture design, API docs).

For each document type, we use specialized chunking strategies. For example resume documents guarantee project experience completeness, QA documents maintain QA pair integrity, and technical documents preserve code blocks and tables.

Innovation 2: Project Completeness Protection Algorithm

We designed a project completeness protection algorithm with key innovations including Markdown header recognition (automatically identify project name format), smart section header merging (merge section headers into the first project), and dynamic overflow tolerance (allow project chunks to exceed limits by 30%).

This algorithm draws from Google Search's "entity recognition" technology, treating projects as complete entities and ensuring internal fields aren't fragmented.

Innovation 3: Context Tag System

We introduced a context tag system that adds semantic tags to each chunk: project tags (include project name, timeline, role), tech stack tags (automatically extract technology keywords), and achievement tags (identify quantitative metrics).

These tags not only improve recall precision but also provide richer contextual information during responses.

Results: Reaching Industry-Leading Levels

After comprehensive testing, we achieved significant results:

JSON Resume before 100% and after 100%, maintained perfection.

Markdown Resume before 0% and after 93%, improved +93%.

TXT Resume before 100% and after 100%, maintained perfection.

QA Document before 100% and after 100%, maintained perfection.

Additional benefits include average chunks reduced by 30-40%, parsing plus chunking time less than 2ms, and memory usage less than 10KB.

These results have reached industry-leading levels. According to our research, knowledge base products on the market typically achieve 30-50% project completeness for Markdown resumes, while we reached 93%, far exceeding industry average.

User Best Practices: How to Prepare High-Quality Knowledge Base

Technical optimization is only one side of the coin. The other side is how users prepare documents. We summarized best practices from interviewers at major tech companies like ByteDance, Alibaba, and Tencent.

Format Selection Priority

Recommendation priority: JSON (Cloud Resume) over Markdown over TXT over Word over PDF.

JSON is structured data with complete fields. If you use Interview AiBox's cloud resume feature, the exported JSON format gets 100% perfect support.

Based on our analysis of 1000+ resumes, high-quality Markdown resumes follow this structure: use third-level headers for project titles, use list format to describe background solution and result, leave a blank line between each project, and provide quantitative data.

ByteDance interviewer advises: "I value quantitative results most. Don't just say 'optimized performance', say 'reduced API response time from 500ms to 50ms, a 10x improvement'."

Alibaba interviewer advises: "Project descriptions should demonstrate your depth of thinking. Background should explain business value, solution should explain technology trade-offs, results should explain business impact."

Document Quantity Recommendations

Knowledge base document quantity should be no more than 30.

According to our test data, 1-10 documents achieve 95%+ recall precision, 11-20 documents achieve 85-95% recall precision, and 21-30 documents achieve 70-85% recall precision.

Best practices include: resume 1 main resume plus 2-3 resume variants for different positions, QA documents 2-3 high-frequency interview question documents, and technical documents 1-2 core tech stack documents.

topK Parameter Configuration

topK determines how many chunks AI recalls from the knowledge base. Resume documents topK equals 7 to prioritize recalling project experience, QA documents topK equals 5 to keep QA pairs complete, and technical documents topK equals 7 to cover more technical details.

Project experience in resumes is the most important information, requiring more chunks to cover complete project descriptions. QA documents usually have one complete QA pair per chunk, so fewer chunks are needed.

Four Project Formats We Support

To make it easier for users to prepare documents, we support multiple project formats:

Format 1 field-based: project name, project time, role, and project description each on separate lines.

Format 2 Markdown header: use third-level header as project name, list format for content.

Format 3 continuous text: project title followed by continuous project description.

Format 4 mixed format: combination of the above formats.

Performance Optimization: From Milliseconds to Microseconds

We not only improved recall quality but also optimized performance:

Before optimization parsing time about 50ms, chunking time about 30ms, memory usage about 100KB.

After optimization parsing time 0ms (JSON cache), chunking time 1ms, memory usage less than 10KB.

Optimization methods include JSON resume parsing result caching, chunking algorithm optimization, and memory reuse.

These optimizations bring our performance to industry-leading levels, processing a resume in just 1ms, more than 10x faster than industry average.

Continuous Iteration: Deepening Knowledge Base Capabilities

We've reached industry-leading levels, but we won't stop here. We'll continue deepening our knowledge base capabilities with cutting-edge technologies:

Intelligent Content Understanding: Introducing NLP technology to automatically extract key information like technical highlights, business value, and team collaboration from projects, providing AI with richer context.

Multimodal Support: Supporting images, charts, and other non-text content, enabling AI to understand architecture diagrams, flowcharts, and other visual content.

Personalized Recall: Dynamically adjusting recall strategies based on user's interview position and company characteristics, providing more precise content.

Real-time Learning: Continuously optimizing recall models based on user feedback and usage data, making the system smarter with use.

Summary

This optimization made us deeply realize: In AI applications, data quality is more important than algorithms.

No matter how powerful the LLM is, if the recalled knowledge fragments are incomplete, it cannot generate high-quality answers. We increased Markdown resume project completeness from 0% to 93% by refactoring the chunking strategy, reaching industry-leading levels.

Core experience includes: semantic completeness is more important than chunk size, multi-format adaptation is a foundational capability, context tags improve recall precision, and user education is as important as technical optimization.

We invested heavily in building a quality assurance system with extensive test coverage: 17 test cases covering multiple document formats and scenarios, 6 scenario tests validating effectiveness across different use cases, automated testing ensuring quality stability for every optimization, and continuous monitoring of recall quality metrics for quantifiable results.

We hope these optimizations help you better showcase your abilities in interviews. If you encounter issues using the knowledge base, feel free to submit issues on GitHub.

Related Links: