Applying Large Language Models to Zhihu's Bridge Platform: Use Cases, Challenges, and Solutions

Introduction – The Bridge platform is Zhihu's internal operations analysis system. The article shares how large language models (LLMs) are applied to automate knowledge organization, natural‑language filtering, and data analysis, discussing both business impact and technical challenges.

1. Business Status and Background – Bridge supports six core scenarios: finding people, content, monitoring, opportunity discovery, and problem investigation. LLMs are used to combine external news‑driven inspiration with internal content‑driven inspiration, forming the first stage of knowledge taxonomy construction.

2. Knowledge System Classification – Two business forms are addressed: event aggregation from external news and content sedimentation into hierarchical taxonomies. LLMs assist in extracting key information, performing multi‑round clustering, naming clusters, and merging similar events, while a MapReduce‑like pipeline mitigates max‑token limits and workflow complexity. Advantages include automatic event naming and higher accuracy.

Event aggregation pipeline: news vectorization → high‑precision clustering → naming via LLM → hierarchical merging → final event generation.

Knowledge organization pipeline: content splitting → map phase (generate category names) → reduce phase (merge categories) → recursive merging until convergence.

Key solutions for token limits, clustering accuracy, and process complexity involve hierarchical clustering of LLM‑generated events, prompt size reduction, and a MapReduce‑style parallel framework.

3. Natural Language to Filter Conditions – Targets packaging, person, and content search. The task involves many filter criteria and complex logical combinations. The team fine‑tuned LLMs across four data‑construction stages (atomic conditions, logical combinations, fuzzy statements, and error‑prone cases) and iterated three model versions to address JSON truncation, output duplication, and logical errors.

Improvements included expanding token limits, random sampling to avoid repetitive outputs, and extensive JSON sample generation to correct format issues.

4. Natural Language Data Analysis – Focuses on ad‑hoc SQL generation from user queries. Challenges include handling diverse NL inputs, mapping queries to appropriate data sources, and ensuring business‑specific results. The solution uses dynamic prompts with FAISS‑based similarity search (MMR) to retrieve relevant examples, construct concise prompts respecting token limits, and generate SQL statements.

Online results show reduced cost, higher user adoption, and streamlined onboarding for new users, though accuracy remains a concern for complex queries, prompting future fine‑tuning efforts.

5. Summary and Outlook – The experience highlights pain points such as lack of mature PE practices, max‑token constraints, prompt engineering trial‑and‑error, model latency, and insufficient frameworks for large‑scale scenarios. Future directions include building dedicated frameworks for complex LLM tasks, leveraging business imagination to expand model capabilities, and continued fine‑tuning.

6. Q&A – Addresses model selection for event aggregation, evaluation methods, termination criteria for knowledge organization, and practical tips for improving LLM performance in production.