Project BaixiaoSheng: An AI‑Powered Project Management Assistant – iQIYI Case Study

From December 5‑7, the 13th TOP 100 Global Software Case Study Summit concluded with the theme “Future‑Oriented Organizational Evolution and Innovation Management”. Over 100 senior R&D leaders and industry experts gathered to discuss how organizations can adapt and transform in the era of large language models (LLMs).

At the summit, iQIYI’s Project Management Office (PMO) shared the practical experience of applying AI to project management through an internal assistant called Project BaixiaoSheng . The assistant integrates static knowledge Q&A, dynamic data consulting, and scenario‑specific assistance to address the growing complexity of parallel projects.

Key Challenges Identified

Multiple parallel tasks leading to frequent context switches and manual effort.

Information overload: abundant unstructured documentation makes knowledge retrieval difficult.

High data‑flow cost: semantic‑level data exchange across tools is cumbersome and time‑consuming.

Solution Architecture

Project BaixiaoSheng is built on a Retrieval‑Augmented Generation (RAG) framework and consists of three core functions:

Static Knowledge Q&A : A curated knowledge base powered by a generic RAG pipeline. Challenges such as “knowledge curse”, ambiguous terminology, and proprietary acronyms are mitigated through synonym normalization, prompt engineering, and multi‑stage retrieval (vector + tag recall, deduplication, rerank).

Dynamic Data Consulting : A plugin‑based system that bypasses the RAG retriever when up‑to‑date data is required. User intent and entities are recognized, then appropriate plugins (including composite plugins that orchestrate multiple calls) fetch real‑time information, which is fed back to the LLM for answer generation.

Scenario Assistant : Extends the dynamic consulting flow to trigger actions in external tools (e.g., creating Feishu groups, annotating requirements, extracting group emails). This bridges conversational queries with concrete workflow automation.

Technical details include:

Synonym unification (e.g., mapping “VIP” to “member”) during both indexing and retrieval.

Query optimization via LLM‑driven completion and historical context enrichment.

Hybrid retrieval using vector similarity and tag matching, followed by deduplication and reranking.

Plugin framework with unique IDs, descriptions, and a composite plugin mechanism for multi‑step data gathering.

Impact

Since launch, the static Q&A module provides 24/7 real‑time responses, improving project transparency and reducing communication bottlenecks. Dynamic data consulting enables users to obtain up‑to‑date information without leaving their chat client, and the scenario assistant automates repetitive cross‑system tasks, significantly boosting efficiency.

Future Roadmap

Training a dedicated LLM fine‑tuned on iQIYI’s proprietary knowledge and policies.

Optimizing the knowledge base with multi‑model fusion and adaptive model selection.

Enhancing recall strategies with richer rerank mechanisms and feedback‑driven threshold adjustment.

Expanding data analysis and decision‑support visualizations.

Developing AI‑Agents to orchestrate end‑to‑end business processes.

These planned upgrades aim to enlarge the internal knowledge scope, improve precision, and broaden the assistant’s applicability across the organization.