Biopharma investment research system / 02
Biopharma Investment Research Agent & Knowledge Base
An agent workflow and knowledge base for biopharma investment research. Public research materials are organised into traceable layers of source texts, summaries and classifications. Agents handle batch parsing, summarisation and classification; I define the processing rules, design the taxonomy and review results. Further knowledge extraction and investment research applications remain in development.
Results & evidence
Parsing and summarisation are complete for 5,819 items, including 129 videos totalling 136.9 hours, yielding approximately 294 million characters. The classification system developed from these summaries covers 47 object types, 39 event types, 52 relation types and 129 research topics, with rules for identifying and merging records and retaining their sources. A model gateway and scheduler built for the project ran 256 concurrent tasks per batch, with peak throughput of approximately 18,000 tokens/s.
Parsing and summarisation completed for 5,819 items
Material counts by format and progress through parsing, summarisation and classification.
View full figure ↗Objects and events associated with a drug in development
An example centred on a drug in development shows the relationships between objects and events. Labels below each node identify its type, while amber dots mark associated events. Six linking rules appear below. Vocabulary v2.1 defines 47 object types, 39 event types and 52 relation types; the figure shows a subset.
View full figure ↗Record & scope
The project applies the layered storage and source-tracing methods used in the Quant Research Library to Chinese innovative medicines and biopharma investment research.
Evaluation settings & scope
Figures are current as of 29 October 2026. The corpus consists of publicly available biopharma research materials. A further 96 reference items, including data tables, are stored separately for lookup.
The summaries have not been individually verified by a human. Their numerical claims were separately checked for corresponding content in the parsed text. They also contain 86,688 candidate relationships between objects, retained for later extraction and not yet counted as knowledge-base entries.
The classification system has passed structural checks. Semantic coverage has not yet been measured against the 95% target. Item-level extraction and retrieval remain in progress, and their results are not presented here.