Knowledge Base Retrieval and Recall for Live Attenuated

Data Characteristics

Pharmacovigilance data for live attenuated vaccines primarily originates from clinical trial reports, real-world evidence (RWE) data, vaccine adverse event monitoring systems (e.g., VAERS, EudraVigilance), and literature reviews. Data updates typically occur quarterly or semi-annually, with urgent updates triggered by significant safety signals. Document structures are predominantly semi-structured and unstructured, including Case Report Forms (CRFs), medical text records, patient interview records, and laboratory test results. Key fields involve vaccine batch numbers, vaccination dates, Adverse Event (AE) names, severity, outcomes, causality assessments, medical history, and concomitant medications. Units for dosage usually appear as titers or International Units (IU). Time is measured in days, months, or years. Laboratory indicators use various units depending on the specific testing method.

Constraints on Knowledge Base Retrieval and Recall

The complexity of live attenuated vaccine data sources requires the knowledge base to effectively integrate multi-source heterogeneous data. This directly impacts document preprocessing and chunking strategies. Irregular update frequencies mean the knowledge base must support incremental updates and version management to ensure timely retrieval results. Semi-structured and unstructured document formats, such as free-text descriptions of adverse events in medical records, pose challenges for knowledge segmentation granularity and semantic understanding. This can lead to critical information being truncated or losing context during segmentation. Furthermore, the diversity of fields and units, especially in adverse event descriptions, necessitates more refined entity recognition and standardization. Without this, issues like "inconsistent symptom names" or "dosage unit confusion" can arise, affecting retrieval accuracy and recall rates.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk size`	`500–800 characters`	Balances the completeness of adverse event descriptions with retrieval efficiency, preventing overly long segments from introducing irrelevant information.
`Chunk Overlap Length`	`100–150 characters`	Ensures continuity of critical medical entities and context across segments, reducing semantic loss due to segmentation truncation.
`Recall count`	`Top 8–12 entries`	Filters the most relevant high-quality recall segments related to live attenuated vaccine adverse events while ensuring coverage.
`Similarity threshold`	`Calibrate by actual measurement`	Requires testing against different adverse event types and query patterns to ensure recall results are neither too broad nor too narrow.
`Rerank result count`	`Top 5 entries`	Re-ranks initial recall results to further improve the precision of the information ultimately presented to the user.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Accommodates the parsing time for large clinical trial reports or complex case report forms, preventing file parsing timeouts.

Common Pitfalls

The knowledge base fails to activate during chat conversations but works correctly in debug preview, showing a "knowledge base not selected" prompt. This occurs when the dataset_ids parameter is not correctly passed or is empty during external API calls, preventing the request from associating with the specified knowledge base.
The knowledge base output cites incorrect files, where the cited source does not match the actual answer. This happens due to an unreasonable document segmentation strategy, leading to critical information being split across multiple unrelated segments. The model may then cite semantically mismatched segments when generating answers.
Retrieval results contain a large amount of irrelevant vaccine product information, indicating low relevance in recalled items. This occurs when the knowledge base index does not effectively differentiate between different vaccine types or product batches, leading to over-generalization during retrieval.

Verification Steps

Select queries for live attenuated vaccine adverse events of varying severity and types. Check if the number of recalled items consistently falls within the configured range.
For queries including specific vaccine batch numbers or rare adverse events, verify that the recall results precisely locate relevant documents and their specific segments.
Examine the similarity threshold distribution of retrieval results. Confirm that most valid recalled items have similarity values within the expected range. Adjust Similarity threshold accordingly.

The values provided are common starting points. Measure them 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.