DataLeap "Find Data Assistant": Leveraging Large Language Models for Data Asset Retrieval and Management

In the digital era, data has become a critical asset, but the explosive growth of data volumes makes efficient discovery and utilization a major challenge; the DataLeap team responded by creating the "Find Data Assistant" that integrates large language model (LLM) capabilities into a data‑asset management platform.

The platform tackles three core problems: (1) rapid location of target data within massive datasets, (2) accurate interpretation of user intent beyond simple keyword matching, and (3) providing concise, context‑aware answers rather than raw document lists. Traditional keyword search suffers from semantic blindness, loss of context, synonym handling issues, and low accuracy.

To overcome these limitations, a hybrid architecture was designed that combines keyword retrieval, semantic (vector) retrieval, and LLM‑driven processing. The workflow proceeds from a user query through a dialogue framework, intent and entity recognition (via LLM), parameter assembly, retrieval from three storage back‑ends (Vector DB for embeddings, Elasticsearch for term matching, MySQL for conversation memory), coarse‑to‑fine ranking, and finally LLM‑based mixed ranking and answer summarization.

Key technical enhancements include optimizing embedding models, introducing a reranker for relevance sorting, employing LLMs for mixed ranking to improve semantic understanding and cross‑domain fusion, and mitigating token‑limit and hallucination issues through large‑token models, micro‑tuning, and streaming output to reduce latency.

Beyond retrieval, LLMs are used for answer summarization, refusal handling, and automatic FAQ generation, enabling continuous enrichment of the knowledge base and reducing manual effort in knowledge‑management workflows.

Lessons learned emphasize trusting LLM capabilities while recognizing the need for small‑model optimization, extensive micro‑tuning, and future integration of agents to handle multi‑turn dialogues and ambiguous intents, guiding ongoing improvements to the system’s performance and reliability.