Reference and Traceability for Internal Office Assistants in

Data Characteristics

Policy retrieval data in the biopharmaceutical industry primarily comes from internal company regulations, operational SOPs, quality management system documents, R&D process specifications, compliance requirements, and the latest national and industry regulations. These documents typically exist as PDFs, Word files, or internal knowledge base pages (e.g., Confluence, SharePoint). Update frequency varies: core policy documents are usually revised annually or biennially, while SOPs for specific projects or experiments may update frequently with project progress. Document structure typically includes fields like title, chapter, clause number, revision history, effective date, and issuing department. The primary language is Chinese, but some R&D documents or international collaboration files may include English. Units of measurement, involving experimental data and production batches, are highly standardized.

Constraints on "Reference and Traceability"

The authoritative nature of policy documents requires precise referencing to the original text location to ensure information credibility. Frequent revisions make version management crucial; references must specify a particular version to avoid confusion. Document hierarchical structures and clause numbers, such as "Quality Management System Document - QP-001 - Version 02 - Clause 3.2.1", require the referencing mechanism to parse and display such structured information. Multilingual content, especially mixed Chinese and English, challenges text segmentation and embedding models, requiring accurate semantic recognition across languages. Additionally, when policies involve sensitive information or access control, the traceability mechanism must integrate with enterprise internal permission systems to ensure only authorized users can view original documents.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`Chunk size` (Segment Length)	`500–800 characters` (characters)	Policy clauses are typically logically complete. Shorter segments may split semantics, while longer ones add irrelevant information.
`Chunk Overlap Length` (Segment Overlap Length)	`50–100 characters` (characters)	Ensures contextual continuity, especially when clauses span pages or paragraphs.
`Recall count` (Recall Count)	`Top 5–8 entries` (top 5–8 items)	Policy retrieval often requires high precision. Increasing recall count appropriately covers potentially relevant clauses.
`Similarity threshold` (Similarity Threshold)	`0.75–0.85`	Policy texts are highly specialized, requiring high semantic similarity to prevent generalized matching.
`Rerank result count` (Reranked Return Count)	`Top 3 entries` (top 3 items)	Performs a secondary sort on the recalled items to focus on the most relevant policy clauses.
`Reference Link Format`	`Original Link + Version Number + Page Number/Clause Number`	Meets strict policy traceability requirements by specifying the exact revision version and location.

Common Pitfalls

Retrieval results fail to cite English documents in the knowledge base. This occurs when the tokenizer or embedding model is not optimized for multilingual mixed texts, leading to insufficient semantic understanding of English content.
Returned reference source links are broken or point to incorrect versions. This usually happens when the knowledge base connector does not correctly handle dynamic links or version management mechanisms of internal document systems.
After a user query, the assistant provides a clear answer but lacks specific policy clause numbers. This is due to an overly coarse segmentation strategy that fails to retain clause numbers as important metadata and embed them with the text.

Configuration Verification

Select several representative policy documents, including mixed Chinese and English documents. Simulate user queries and check if references accurately point to specific clauses, page numbers, and version numbers in the original text.
View the returned reference list via API or interface to confirm that the Reference Link Format meets expectations and successfully accesses the original document.
Test with policy documents containing revision history. Verify if the assistant prioritizes citing the latest effective version in its answers and can indicate historical version information when necessary.

Note: 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.