Technical Highlights of an Enterprise RAG Knowledge Base System

Purpose of the document: pre-sales solution material for an enterprise RAG system PPT
Core techniques: LLM-based information organization + multi-level chunk strategy
Applies to: pre-sales solution knowledge bases, multi-source knowledge asset consolidation inside enterprises

1. Core challenges and our approach

Pain points of traditional RAG

In enterprise pre-sales scenarios, a knowledge base faces three major challenges:

Our approach: Context Engineering

Core idea: the key to RAG is not "retrieval + generation", but deciding, at every LLM generation step, what information should enter the context window.

Through "knowledge assetization" we turn messy raw material into "digital assets" that AI can call efficiently, leaping from "searching documents" to "professional answers in seconds".

2. Highlight 1: an LLM-driven information organization pipeline

2.1 Overall pipeline architecture

Raw PPT/PDF -> extract -> per-page summary -> LLM preprocessing -> layered aggregation -> vector ingestion
     ↓              ↓           ↓               ↓               ↓
  multimodal    text/image    structured       Category        high-density
                            understanding     aggregation        chunks

2.2 The core value of LLM preprocessing

Before information is written into the vector store, we introduce an LLM preprocessing layer to deeply organize the raw information:

① Automatic Q&A pair generation (QAGenerator)

Input: the content summary of a single PPT/PDF page
Processing: the LLM generates 5-10 high-quality Q&A pairs from the page content
Output: structured QAPairs (question, answer, keywords, confidence)

Technical highlights:

• simulates the high-frequency question patterns of pre-sales scenarios
• generates paraphrases (Alt Questions) to improve recall coverage
• confidence scoring filters out low-quality content

② Semantic classification and aggregation (LLMClassifier)

Input: the full list of Q&A pairs
Processing: batch classification into predefined business Categories (8 major classes)
Output: Q&A groups aggregated by category

Technical highlights:

• batching cuts classification cost by 80% (batch_size=8)
• a hybrid strategy with rule-based classification balances accuracy and efficiency
• falls back to heuristic classification when the LLM fails

③ Topic organization and summary synthesis

Grouping stage (local analysis):

• send Q&A pairs of the same category to the LLM in batches
• generate the local summary and representative Q&A of that topic

Topic synthesis stage (global synthesis):

• merge all local summaries
• have the LLM rewrite them into an authoritative summary with a single voice
• attach 3-5 high-value typical Q&A pairs

Technical highlights:

• restructures linear, verbose raw text into structured, high-density knowledge fragments
• collapses and clusters "information entropy" (deduplication, classification, convergence)
• resolves the contradiction between macro user questions and micro data

2.3 Quality assurance before ingestion

3. Highlight 2: layered chunk strategy (RAPTOR-style)

3.1 Three-layer knowledge architecture

We adopt a Summary-first layered retrieval strategy and split knowledge into three levels:

+-----------------------------------------------------+
|  L1: Overview (project profile layer)               |
|  "what it is", "what it can do"                     |
|  Use: project navigation, high-level overview,      |
|       global context                                |
+-----------------------------------------------------+
|  L2: Category Summary (category summary layer)      |
|       ★ primary retrieval unit                      |
|  "how it works", "the details"                      |
|  Use: the workhorse for professional questions,     |
|       high-density stable recall                    |
+-----------------------------------------------------+
|  L3: Slide Evidence (raw evidence layer)            |
|  "why should I believe this"                        |
|  Use: evidence supplement, citation support,         |
|       anti-hallucination verification               |
+-----------------------------------------------------+

3.2 Chunk type design (six types)

3.3 Layered retrieval strategy

Two-stage retrieval

First hop (high-density layer):

• restrict the search space: chunk_type in {overview, category_summary}
• hit Category Summary first (the primary retrieval unit)
• return a small number (3-5) of high-quality summaries

Second hop (evidence layer):

• extract source page numbers from the hit results' source_slide_refs
• retrieve Slide Evidence on demand to fill in detail
• used by CRAG (corrective retrieval) for evidence verification

Technical advantages:

• avoids blind search across a huge pile of fragments
• each layer returns content of controlled length, so the context window is used efficiently
• supports a CRAG-style corrective loop

3.4 Chunk standardization

The unified ChunkDocument structure

{
  "id": "ChatBI_category_capability",
  "text": "【Category】核心能力...",
  "metadata": {
    "project_name": "ChatBI",
    "chunk_type": "category_summary",
    "level": "category",
    "category_id": "capability",
    "category_name": "核心能力",
    "source_slide_refs": [3, 4, 5, 8],
    "source_files": ["page_summaries/003.json", ...],
    "confidence": 0.92,
    "rag_schema_version": "v2"
  }
}

Standardized ID naming rules

4. Highlight 3: a configurable policy switch system

4.1 Flexible policy combinations

The system provides fine-grained feature switches to fit different cost budgets and accuracy requirements:

# 核心功能开关
enable_category_summary_chunks: true    # 类别摘要（主力，默认开）
enable_llm_classify: true               # LLM分类（vs 规则分类）

# 可选增强（默认关，按需启用）
enable_qa_pair_chunks: false            # 单页问答对
enable_metrics_chunks: false            # 指标类Chunk
enable_topic_chunks: false              # 主题类Chunk（legacy）
enable_step_chunks: false               # 流程类Chunk（legacy）

4.2 Dual-mode operation

Technical details of automatic mode:

• LangGraph orchestrates the topic organization and summary synthesis state machine
• nodes: batch preparation → category organization → source merge → topic synthesis → overview construction → validation
• supports concurrency control, failure fallback, and mock tests

5. Highlight 4: alignment with mature paradigms

Our solution aligns closely with mature RAG architectures in academia and industry:

6. Implementation advice and best practices

6.1 Recommended configuration

Pre-sales solution knowledge base (recommended configuration):

# 主力Chunk
enable_category_summary_chunks: true
enable_overview_chunk: true

# 分类策略
enable_llm_classify: true               # 高准确率

# 可选补充（视预算开启）
enable_qa_pair_chunks: true             # 补充召回
enable_metrics_chunks: true             # 性能指标问答

# 关闭低ROI功能
enable_topic_chunks: false
enable_step_chunks: false

# 检索参数
top_k: 8                                # 召回候选数
top_n: 5                                # 最终选取数
tau: 0.5                                # 相似度阈值

6.2 Knowledge base expansion paths

The current pre-sales solution knowledge base can serve as the first landing scenario of an enterprise RAG platform, and can later expand to:

1. Technical documentation: API docs, technical specifications → use step + topic chunks
2. Training material: courses, lecture notes → use qa_pair + overview chunks
3. Case library: project cases, customer feedback → custom chunk types

Unified access: all knowledge bases follow the same chunk standard contract, so they can share a retrieval engine and a QA system.

7. Summary

Core value proposition

1. Information organization comes first: before writing to the vector store, an LLM completes the full "understand → classify → aggregate → summarize" pipeline
2. Layered retrieval architecture: a RAPTOR-style three-layer chunk design aligns with the cognitive path of "overview → detail → evidence"
3. Traceable and auditable: every knowledge unit carries complete source traceability, enabling "every answer is an audit"
4. Flexible and controllable: fine-grained switches plus dual-mode operation fit different cost budgets and quality requirements

Ideal customer profile

• enterprises holding large volumes of unstructured solution material (PPT/PDF/Word)
• B2B scenarios with high demands for answer accuracy and traceability
• organizations that need a unified knowledge platform reused across business lines

Appendix: key term glossary