Knowledge Base Retrieval and Recall for Bispecific Antibody

Data Characteristics

Bispecific antibody pharmacovigilance data primarily originates from clinical trial reports, real-world evidence (RWE) studies, post-market surveillance reports, medical literature, and regulatory agency databases (e.g., FDA Adverse Event Reporting System, FAERS). This data updates frequently, especially early after drug approval, with new reports potentially appearing weekly or even daily. Document structures are diverse, including structured Case Report Forms (CRF), semi-structured medical texts (e.g., physician's handwritten notes, summaries), and unstructured free-text descriptions. Data fields cover patient demographics, medication history, comorbidities, adverse event (AE) descriptions, severity, onset time, outcome, drug dosage, administration route, and duration. Adverse event descriptions often include medical terminology (e.g., MedDRA codes) and natural language. Units involve dosage (mg/kg), time (days, weeks), and frequency (times/day).

Constraints Imposed on Knowledge Base Retrieval and Recall

The high update frequency of bispecific antibody data requires the knowledge base to support rapid incremental updates and indexing, ensuring the timeliness of retrieval results. The diversity of document structures means the knowledge base must ingest and parse various data formats. This includes extracting key fields from structured data and effectively segmenting and vectorizing unstructured text. The medical specificity of adverse event descriptions, particularly the co-existence of MedDRA codes and natural language, demands robust semantic understanding and query expansion to avoid recall omissions due to terminology mismatches. Patient variability and complex medication backgrounds mean that simple keyword searches are insufficient to capture multi-dimensional associated information, requiring support for multi-field co-queries or context-aware retrieval. Time-related aspects (e.g., AE occurrence relative to drug administration window) also require the retrieval model to process time-series information and identify potential temporal patterns.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale for Recommendation
`Chunk size` (Segment Length)	`300–500 characters` (characters)	Balances the completeness of adverse event descriptions with vectorization efficiency, preventing dilution of key information by overly long texts.
`Chunk Overlap Length` (Segment Overlap Length)	`50–80 characters` (characters)	Ensures contextual continuity, especially in medical event descriptions where adjacent sentences have strong relevance.
`Recall count` (Recall Count)	`Top 10–20 entries` (top 10–20 items)	Given the complexity and diversity of bispecific antibody adverse reactions, increasing recall quantity improves coverage.
`Similarity threshold` (Similarity Threshold)	`Calibrate by actual measurement` (calibrate based on actual measurements)	Requires adjustment based on the specific embedding model and dataset to balance precision and recall.
`Rerank result count` (Rerank Return Count)	`Top 5 entries` (top 5 items)	Further refines initial recall results through a reranking model, focusing on the most relevant outcomes.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds` (seconds)	Allows sufficient parsing time for large clinical reports or summary documents, preventing timeouts.

Common Pitfalls

Retrieval results contain a large amount of irrelevant or low-relevance adverse event information. This happens when the Similarity threshold (Similarity Threshold) is set too low, leading to noise interference.
Some critical adverse event reports are not recalled, resulting in incomplete query results. This occurs when Chunk size (Segment Length) is too long, diluting key information, or Recall count (Recall Count) is too small, failing to cover potentially relevant documents.
After a knowledge base update, retrieval results still show old information. This indicates that the incremental indexing mechanism was not triggered correctly, or the index update frequency configuration does not match the data source's update speed.

Verification Steps

For typical adverse event queries, cross-reference recall results with known important literature and reports, then evaluate their ranking.
Conduct multi-turn dialogue tests to observe if the system accurately understands and recalls detailed data related to specific bispecific antibody adverse reactions based on contextual information.
Simulate new adverse event reports to verify if the knowledge base's incremental update mechanism incorporates new data into the retrieval scope within the configured time.
Use mixed queries containing both medical professional terminology and natural language descriptions to assess the retrieval system's compatibility and accuracy with different query forms.

The values provided are common starting points and should be measured against the reader's 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.