Knowledge Base Retrieval and Recall for Small Molecule Drug

Data Characteristics

Data for small molecule drug clinical trial prescreening comes from drug compound databases (e.g., PubChem, ChEMBL), patent literature, drug development reports, preclinical study data, and published scientific papers. Data update frequencies vary. Compound structure and physicochemical property data are relatively stable. Preclinical data and patent information are dynamic, with new batches updated monthly or quarterly. Document structures typically include fields for structural formulas (SMILES, InChI), molecular weight, LogP, solubility, and other physicochemical properties, as well as target sites, pharmacodynamic data, toxicology reports, pharmacokinetic parameters, and preclinical trial protocol descriptions. Field units are strict. For example, molecular weight is in Da, concentration in nM or µM, and dosage in mg/kg.

Constraints on Knowledge Base Retrieval and Recall

The specific nature of small molecule drug data imposes particular requirements on knowledge base retrieval and recall. Structural formula information necessitates support for structure-based similarity retrieval; traditional text matching is insufficient to capture chemical semantics. The numerical nature of physicochemical property fields requires the knowledge base to support numerical range queries and unit conversions for precise retrieval. Preclinical trial protocol descriptions often contain extensive experimental details and specialized terminology. Their semi-structured text characteristics make tokenization and entity recognition crucial. Inconsistent data update frequencies require the knowledge base to support incremental updates and version management to avoid recalling outdated or duplicate information. Furthermore, different data sources may have subtle differences or even conflicts in descriptions of the same compound or target. The recall mechanism needs to identify and flag potential data inconsistencies.

Configuration Settings

Configuration Item	Recommended Value	Rationale for this Value
`Chunk size` (Segment Length)	`500–800 characters`	Ensures a single segment can contain complete physicochemical properties of a compound or experimental procedure descriptions, preventing semantic fragmentation.
`Recall count` (Number of Retrieved Items)	`10–20 items`	Balances recall coverage with the processing efficiency of subsequent reranking modules, avoiding excessive redundancy.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Requires a higher threshold for specialized terminology and structured data in small molecule drugs to filter out irrelevant results.
`Rerank result count` (Number of Reranked Items)	`3–5 items`	Focuses on the core information most relevant to the query, reducing model processing load and improving response speed.
`embeddingModel`	`bge-large-zh-v1.5`	Better understands specialized vocabulary and contextual semantics in the biomedical field.
`maxContext`	`4000 characters`	Accommodates the contextual needs of long texts like clinical trial protocols, ensuring the model receives sufficient information.

Three Common Mistakes

Symptom: The retrieval results contain a large number of irrelevant compounds or target information, not matching the query intent. Reason: The Similarity threshold (Similarity Threshold) is set too low, failing to effectively filter out low-relevance document segments.
Symptom: The knowledge base is referenced, but the generated answer does not fully utilize the specialized knowledge base content, instead mixing in general model responses. Reason: Recall count (Number of Retrieved Items) or Rerank result count (Number of Reranked Items) are set incorrectly, causing the model to prioritize high-quality recalled results from the specialized knowledge base within a limited context window.
Symptom: Relevant data clearly exists in the knowledge base, but it cannot be retrieved, or the retrieved content is incomplete. Reason: The Chunk size (Segment Length) is set too short, causing critical compound structures, physicochemical properties, or experimental method descriptions to be split into different knowledge blocks, affecting semantic completeness.

How to Confirm Correct Configuration

For typical small molecule drug queries, check if the recall results accurately point to relevant compounds, targets, or experimental data, and verify the correctness of units and values for key fields.
Randomly select multiple documents from the knowledge base, simulate queries for their core content, and observe the completeness and relevance of the retrieved items, especially documents containing structural formulas or complex experimental protocols.
By comparing the quality and quantity of recall results at different Similarity threshold (Similarity Thresholds), determine a range that effectively balances recall and precision.

The values provided are common starting points and should be measured against your own samples.

Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-21.