Multi-Turn Conversations and Prompts for Gene Therapy AAV

Data Characteristics

Gene therapy AAV (adeno-associated virus) quality documents include production batch records, test reports, stability study reports, raw material release records, deviation investigation reports, change control documents, and regulatory submission materials. These documents are typically in PDF, Word, or Excel formats. Some data may reside in LIMS (Laboratory Information Management System) or MES (Manufacturing Execution System) export files. Data update frequency correlates with batch production, testing cycles, and regulatory requirements, usually monthly or quarterly. Important documents, such as annual product quality reviews, are updated annually. Document structures are rigorous, adhering to GMP/GLP guidelines, and contain extensive specialized terminology, abbreviations, charts, and data tables. Fields and units are highly specific, for example, titer (vg/mL), purity (%), host cell DNA residue (pg/mL), and endotoxin (EU/mL). Descriptions of test methods and limit requirements often accompany these.

Constraints on Multi-Turn Conversations and Prompts

The specialized and data-intensive nature of gene therapy AAV quality documents imposes specific requirements on the accuracy of multi-turn conversations and prompt construction. The extensive specialized terminology and abbreviations in the documents require the model to possess a high level of domain knowledge to prevent ambiguity or misinterpretation of critical information in multi-turn conversations. The periodicity of data updates means the knowledge base needs regular synchronization with the latest batch data to ensure the timeliness of conversation results. The rigorous document structure and compliance requirements necessitate prompt designs that guide the model to cite specific document sections, batch numbers, or test items in its answers to support traceability. The specificity of fields and units requires prompts to precisely locate and extract numerical values, avoiding unit confusion or misreading of values, especially when calculations or limit comparisons are involved. For example, when asking "Does the titer of a certain batch of AAV product meet the release standard?", the model must accurately identify the batch, extract the titer value, compare it against the standard, and provide a conclusion. This relies heavily on refined prompt guidance and context management.

Configuration Settings

Configuration Item	Recommended Value	Rationale
`maxContext`	`6`	Ensures conversation history covers common multi-turn follow-up scenarios for AAV product quality queries while avoiding irrelevant information interference.
`Chunk size` (Segment Length)	`800–1200 characters`	Accommodates the characteristic of AAV documents having many long paragraphs while maintaining semantic integrity.
`Recall count` (Recall Count)	`Top 10 entries`	Covers relevant information in AAV quality documents that may be dispersed across multiple sections or tables, improving recall rate.
`Similarity threshold` (Similarity Threshold)	`0.75`	Filters out document segments with low relevance to AAV quality queries while ensuring relevance.
`Rerank result count` (Rerank Return Count)	`Top 5 entries`	Focuses on core information most relevant to AAV quality issues, improving answer precision.
`System Prompt`	Calibrate by testing	Must include AAV domain-specific terminology, GMP compliance requirements, and explicitly state that answers should cite document sources.

Common Pitfalls

Numerical misinterpretation or unit confusion in conversations, such as confusing titer (vg/mL) with purity (%). This occurs when prompts do not sufficiently emphasize unit identification or when the model's ability to recognize specific units is insufficient.
The model fails to maintain context for specific batches or test items in multi-turn conversations, leading to off-topic answers in subsequent questions. This occurs when maxContext is set too low or the System Prompt does not effectively guide the model for context tracking.
LaTeX formulas display correctly in debug previews but show raw code after deployment. This occurs when the frontend rendering environment lacks LaTeX parsing library support, requiring frontend integration of renderers like MathJax or KaTeX.

Verification Steps

Test whether multi-turn conversations can accurately track context and provide consistent answers for complex queries involving different batches and test items.
Verify that the model can accurately cite key information from documents, such as batch numbers, test methods, and limit requirements, and trace them back to specific document segments.
Check that for questions involving numerical comparisons (e.g., "Does it meet the standard?"), the model can correctly extract values, units, perform logical judgments, and provide the correct conclusion.
Confirm that the frontend conversation interface can correctly parse and display common LaTeX format formulas or special symbols found in AAV quality documents.

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.