From Knowledge Islands to Marketing Brain: Knowledge Fusion & Graph Reasoning for Corporate Products

Policy Background: The Era of AI-Driven Corporate Marketing

Recent government policies—including the State Council's "Opinions on Deepening the Implementation of the 'AI+' Action" (Guo Fa [2025] No. 11), the NFRA's "Implementation Plan for High-Quality Development of Digital Finance in Banking and Insurance" (Jin Ban Fa [2025] No. 95), the PBOC/NDRC/NFRA/CSRC "Implementation Opinions on 'AI+Finance'" (Yin Fa [2026] No. 101), and the NFRA's "Guidance on Safe Development and Application of AI in Banking and Insurance" (Jin Fa [2026] No. 8)—mandate accelerated AI adoption in finance. China Postal Savings Bank's "15th Five-Year Plan" aligns with this, targeting digital transformation across marketing, service, risk control, and operations. Intelligent product recommendation is a key AI scenario; however, most banks still rely on keyword search or simple rule matching, failing to meet complex corporate banking needs. The core breakthrough is building an understandable, reasonable, explainable knowledge base that converts LLM capabilities into expert-level marketing productivity.

Pain Points: The "Knowledge Island" Dilemma

China Postal Savings Bank manages 200+ corporate products. Knowledge is scattered across structured product fields in management systems and unstructured documents (PDF, PPT, Word, Excel) from seven departments—management measures, operating procedures, application materials, typical cases, marketing scripts, promotional materials—lacking a unified standard. Frontline relationship managers face four core difficulties:

Multi-source heterogeneous fusion difficulty: Structured fields and unstructured long documents have diverse sources, formats, and semantic standards, making unified modeling, alignment, and association hard.

Product associations unavailable: Missing combination, mutual exclusion, affiliation, and scenario relationships between products; traditional keyword and single-table retrieval only achieve literal matching, unable to reflect business logic.

Inaccurate product positioning: User needs cannot be stably mapped to suitable products; cross-department complex needs are even harder to combine into solutions.

Incomplete eligibility queries and unexplainable results: Conditions, materials, and processes are scattered across documents, difficult to retrieve precisely and verify for compliance; direct LLM generation tends to hallucinate, lacking business evidence.

The root cause is "unusable knowledge and relationships." The solution: transform multi-source heterogeneous unstructured knowledge into retrievable, associable, reason-able, explainable structured knowledge assets.

Knowledge Reconstruction: Four-Stage Engineering Practice

The bank built a "Knowledge Precipitation → Knowledge Extraction → Knowledge Storage → Knowledge Application" engineering system (see architecture diagram), integrating 189 corporate products and 11 knowledge categories into a full knowledge asset pool.

1. Knowledge Precipitation: Building a Full Product Knowledge Asset Pool

Structured product data is pulled from the internal product information management system. Business materials, experience summaries, and excellent cases from seven corporate banking departments are aggregated into a "Corporate Product Knowledge Space" with standardized document upload, version management, and access control, achieving centralized governance of multi-source heterogeneous product data.

2. Knowledge Extraction: Dual-Dimensional Tagging Platform

For unstructured-to-structured conversion, a document-level and fragment-level dual tagging system was built. Document-level uses 11 standardized tags (management measures, product introduction, handling process, typical cases, etc.) for product association and type classification. Fragment-level uses seven major categories with 39 sub-tags (basic information, product attributes, customer profiles, structure and returns, associated products and combination logic, marketing handling guides) for paragraph-level fine-grained extraction, yielding a 40+ dimensional product tag library. The platform integrates OCR, five-level text splitting, LLM NLP, and rule reasoning to automatically complete heterogeneous data structuring, providing standardized data foundations for vectorization and graph construction.

3. Knowledge Storage: Dual-Engine Fusion of Vector Database and Knowledge Graph

The storage layer adopts a "vector database + knowledge graph" dual-engine architecture: vectors handle "semantic similarity," graphs handle "business association," together improving recommendation quality and explainability.

Vector Database: Scene-demand oriented, only semantic units "valuable to the scene" are vectorized. Differentiated expression via field and fragment weight allocation: structured core info (product name, category, overview, service scenarios, service objects, term, bank advantages, recommended scripts) weight 0.9; non-core info (recommendation conditions, customer value, related cases) weight 0.1; unstructured paragraph content weight 0.05. This layered weight design ensures core product features dominate semantic matching, avoiding redundant interference.

Knowledge Graph: Product nodes as core skeleton, connecting product category/community nodes, 11 document-type nodes, and document slice nodes. The graph contains 17,622 nodes, 109,893 relationship edges, and 16 communities. Relationship modeling covers products (189), communities, management measures, typical cases, etc., establishing 16 relation types including combination recommendations (259), product mutual exclusions (61), and product-document slice links (73,066). Structured fields provide stable business anchors; unstructured documents attach via "contains-contained by" relations to the product skeleton, enabling unified query of attributes, content, and relationships. The graph supports multi-hop cross-entity reasoning (e.g., "query all products under a community," "query all regulations related to a product"), upgrading the system from document retrieval to explainable reasoning chains—fundamentally shifting from "finding documents" to "locating products and integrating evidence."

4. Knowledge Self-Evolution: End-to-End Automated Closed Loop

The entire precipitation, extraction, and storage pipeline runs end-to-end automatically after launch, zero manual intervention. Business staff only add a product and upload materials in the product information management platform; the system auto-triggers document parsing, dual-dimensional tagging, vectorization, and graph relationship construction. New knowledge typically enters the graph within minutes and takes effect immediately. This "upload-to-fuse, add-to-use" self-evolution mechanism eliminates the traditional knowledge engineering bottleneck of manual annotation and batch refresh, enabling continuous automatic growth of the knowledge base with zero maintenance cost.

Intelligent Recommendation: Four-Layer Progressive Agent Architecture

The application layer builds an "Intent Recognition → Vector Matching → Graph Retrieval → LLM Polishing" four-layer progressive recommendation flow, each layer collaborating to close the loop from user input to precise recommendation.

Layer 1: Intent Recognition. Distinguishes product inquiry vs. product recommendation intent for scene routing. For inquiries, the system introduces product overview, suitable scenarios, service objects; for recommendations, it matches based on customer pain points or needs. Uses LLM for semantic understanding and intent classification, effectively handling colloquial, vague queries from relationship managers.

Layer 2: Vector Matching. Recalls top candidate products via multi-dimensional weighted similarity, solving keyword search's inability to understand semantic fuzziness. The weight allocation mechanism ensures core product features have higher matching priority in semantic space, precisely mapping user needs to candidate product sets.

Layer 3: Graph Retrieval. Queries multi-dimensional associations of target products: combination recommendations, mutual exclusion rules, management measures, typical cases. Traverses relationships to achieve "find related products → integrate supporting documents" reasoning chains, providing business logic support and compliance verification for recommendations. This is the core differentiator from traditional RAG: graph multi-hop reasoning links isolated product info into complete business evidence chains.

Layer 4: LLM Polishing. Integrates retrieval and graph results; prompt engineering performs information filtering, re-ranking, and script generation. The re-ranking LLM follows a five-step process: "special business rule identification → recommendation demand identification and filtering → eligibility verification → combination recommendation relationship mining → mutual exclusion relationship filtering," producing precise, compliant recommendation plans from candidate sets, outputting structured results covering product introduction, recommendation reasons, risk-return explanations.

Risk Prevention: Rule matching, vector retrieval, and graph query collaborate to guarantee data truthfulness and prevent hallucination. The graph's "white-box" nature gives every recommendation explicit business logic support—explainable and traceable. Mutual exclusion relations enable millisecond-level compliance checks, ensuring recommendations are compliant and feasible.

Landing Results: From Knowledge Base to Marketing Enablement

Model Effectiveness

Rigorous evaluation shows system Q&A accuracy at 90%, a 30-percentage-point lift over single LLM+RAG+prompt baseline (60%). Evaluation uses a four-step method: layered dataset construction, same-set A/B testing, strict judgment, calculation verification. Samples drawn from real business questions (product inquiry, single-product recommendation, combination recommendation); business experts provide standard answers, ensuring reliability and business relevance.

System Integration

The solution forms standardized, reusable product base capabilities, deployed as independent entry points in multiple internal business systems, and embedded via process integration to empower key scenarios: small enterprise marketing report generation, customer analysis reports, opportunity expansion, intelligent ticket Q&A, face-to-face talk reference plan generation. As of September 30, 2026, cumulative output of recommendations and product answers reached 36,092 times.

Business Value

The solution frees relationship managers from tedious product queries and solution writing, drastically shortening new-hire onboarding time, enabling 7×24 real-time professional support. In field sales, managers input customer needs on-site to instantly get matched products, recommendation reasons, and original regulations. In complex scenarios, the system integrates cross-department products for large clients, auto-generating comprehensive marketing plans with combination logic, handling paths, and risk alerts—significantly boosting marketing professionalism and timeliness. This exemplifies the bank's "digital-intelligent" upgrade, empowering corporate "comprehensive" and "refined" operations.

Future Outlook: From Product Recommendation to Marketing Decision Intelligence

As LLMs and agents deepen in banking, corporate product recommendation evolves from "passive response" to "active decision." Three expansion dimensions:

Build Customer Digital Twins: Fuse transaction, credit, supply chain multi-source data for 360° customer insight, upgrading from "manual demand entry" to "active perception of customer signals," auto-identifying explicit and latent needs for high-quality recommendation data foundation.

Upgrade to Agentic GraphRAG Architecture: Introduce multi-agent collaboration; auto-select vector retrieval, keyword retrieval, graph multi-hop, structured query strategies by problem complexity; multi-objective re-ranking fuses fit, compliance, customer value, combination synergy, evidence completeness; outputs inclusion and exclusion reasons—leaping from "similarity recall" to "constrained decision."

Establish Dual Closed-Loop Self-Evolution: Task-level loop uses "Plan-Execute-Check-Correct" with confidence thresholds and iteration limits to control cost and risk. Business-result loop continuously collects exposure, view, adoption, handling, conversion, rejection feedback signals; applies causal uplift and contextual multi-armed bandits to optimize next-best-action strategy—moving recommendation from "one-shot" to "continuous learning."

Overall, multi-source heterogeneous knowledge fusion and graph reasoning provide a solid technical foundation for corporate product intelligent recommendation, solving knowledge islands and initiating a marketing paradigm shift from "experience-driven" to "knowledge-driven." The constructed knowledge engineering system and recommendation architecture are highly transferable, applicable to retail, cross-border, and more business lines, offering a reusable methodology and practical path for the bank's "six transformations" digital-intelligent upgrade.