Knowledge Base Retrieval for Meeting Minutes Internal Office

Data Characteristics

Meeting minutes data in the biopharmaceutical domain typically exists as unstructured text. Content includes project progress, experimental data discussions, clinical trial results, compliance requirements, and market analysis. Data sources are primarily meeting recording transcripts, manual shorthand, or automated meeting system generation. Update frequency is high, with new minutes generated weekly or bi-weekly. Document structure is relatively fixed, containing fields such as meeting time, location, attendees, topics, discussion content, resolutions, and action items. Discussion content often involves specialized terminology, drug names, gene sequences, and clinical indicators, with units like nM, μg/mL, and mg/kg.

Constraints on Knowledge Base Retrieval and Recall

The unstructured nature of meeting minutes requires robust text parsing capabilities from the knowledge base. High update frequency necessitates efficient incremental updates and version management to ensure timely retrieval results. Key fields in the document structure, such as meeting time and attendees, can serve as important metadata for filtering and sorting during retrieval. The density of specialized terminology and units in the biopharmaceutical domain requires the tokenizer and embedding model to accurately understand contextual semantics. This avoids recall accuracy issues caused by improper splitting of professional vocabulary. Precise recall of discussion content, especially regarding specific experimental data and resolutions, is crucial for system effectiveness.

Configuration Settings

Configuration Item	Suggested Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters`	Balances semantic completeness and segment recall efficiency. Avoids information loss or inaccurate recall from segments that are too long or too short.
`Chunk Overlap Rate` (Segment Overlap Rate)	`100–150 characters`	Ensures contextual continuity and improves accuracy of cross-paragraph information recall, especially for continuous discussions in minutes.
`Recall count` (Number of Retrieved Items)	`Top 5–8 items`	Balances retrieval speed and recall coverage, ensuring multiple relevant minute segments are covered.
`Similarity threshold` (Similarity Threshold)	`Calibrate based on actual measurements`	Requires adjustment based on actual data recall performance. Ensures retrieved results are relevant and accurate.
`Rerank result count` (Number of Reranked Items)	`3–5 items`	Refines sorting based on initial recall to improve the relevance of the final results presented to the user.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds`	Addresses potentially long parsing times for large meeting minute files, preventing parsing timeouts.

Common Pitfalls

Symptom: When uploading XLSX formatted meeting minutes, some key question-answer pairs are not recognized. Reason: The XLSX file parser might not be optimized for specific table layouts or merged cells, leading to incomplete data extraction.
Symptom: Retrieval results contain numerous irrelevant meeting minute segments. Reason: The Similarity threshold (Similarity Threshold) is set too low, leading to an overly broad recall range and failure to effectively filter low-relevance content.
Symptom: After a knowledge base update, newly uploaded meeting minute content cannot be retrieved promptly. Reason: Incremental indexing strategies are not configured or are improperly configured, or the knowledge base index rebuilding cycle is too long.

Validation Steps

Select multiple meeting minutes containing specific professional terminology and decision content. Perform retrieval tests and check if the recalled results include this key information.
Simulate user queries, such as "latest clinical progress meeting records for [specific drug]". Verify if the results within the Rerank result count (Number of Reranked Items) are highly relevant.
In the knowledge base management interface, check if the latest uploaded meeting minutes are successfully indexed and retrievable via keywords.
Regularly review the performance of Recall count (Number of Retrieved Items) and Similarity threshold (Similarity Threshold) in actual use. Adjust based on user feedback to achieve desired recall precision and coverage.

The values provided are common starting points. Measure 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.