Data Characteristics for This Category
Data for culture media and consumables primarily originates from supplier product manuals, technical specifications, Safety Data Sheets (SDS), and internal procurement and quality control records. These documents are typically in PDF format. Some data may exist as structured tables (e.g., Excel) or within internal databases. Data updates occur when new products launch, batches change, or regulatory requirements adjust, usually quarterly or semi-annually. Document structures are complex, including fields such as product name, batch number, manufacturer, expiration date, storage conditions, component lists, quality control standards, and application ranges. Component lists may involve chemical formulas, CAS numbers, and concentration units (e.g., g/L, %, mM). Quality control standards include test indicators like pH, osmolality, and sterility.
Constraints Imposed by These Characteristics on Multi-Turn Conversations and Prompts
The diversity and complex fields within culture media and consumable data documents challenge the accuracy of multi-turn conversations. For example, when a user queries the storage conditions for a specific batch, the system must accurately extract information from unstructured text and recognize unit-bearing values like "2-8°C". Chemical formulas and concentration units in component lists require the model to possess domain knowledge to avoid confusion or misinterpretation. The update frequency dictates that the knowledge base requires regular maintenance to ensure the timeliness of conversation content. In multi-turn conversations, users may ask about different attributes of the same product across various turns. Prompt design must guide the model to maintain coherence during context switching and handle detailed user questions about numerical ranges or specific testing methods.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Chunk size (Segment Length) | 800–1200 characters | Balances document information density with model processing length limits, ensuring sufficient context within a single segment. |
Recall count (Recall Count) | Top 8 entries | Covers potentially relevant information for user queries, reducing information loss due to insufficient recall. |
Similarity threshold (Similarity Threshold) | 0.75–0.85 | Balances recall precision and recall rate, preventing interference from irrelevant information while ensuring relevant information is retrieved. |
maxContext | 32000 tokens | Accommodates the cumulative context information in multi-turn conversations, especially when asking follow-up questions about product details. |
Rerank result count (Reranked Return Count) | Top 5 entries | Optimizes the quality of the final answer presented to the user, prioritizing the most relevant and high-quality recall results. |
PARSE_FILE_TIMEOUT_SECONDS | 600 seconds | Handles the parsing time for large PDF product manuals or technical specifications, preventing processing failures due to timeouts. |
Three Common Pitfalls
- Phenomenon: A user asks about a product's batch information, but the conversation result does not display the relevant batch number or expiration date. Reason: The knowledge base's segmentation strategy for the corresponding document does not adequately consider the independence of batch information, leading to critical fields like batch numbers being truncated or mixed with irrelevant content.
- Phenomenon: In a multi-turn conversation, a user asks about the components of a culture medium in one turn, then follows up with a question about the CAS number of a specific component, but the system fails to correctly link them. Reason: The prompt design does not effectively guide the model to track specific entities (e.g., component names) and maintain contextual continuity across multiple turns.
- Phenomenon: The latest product specification sheet has been uploaded, but the conversation still returns old version information. Reason: The knowledge base's update mechanism is not synchronized with the data source's update frequency, or index rebuilding is not timely, leading to outdated knowledge base content.
How to Confirm Proper Configuration
- Select a batch of typical culture media and consumable products. Simulate multi-turn conversations from basic product information to detailed technical parameters, checking the accuracy and coherence of the responses.
- Test queries containing special characters, chemical formulas, or specific units (e.g., μg/mL, mOsm/kg) to verify the system's ability to correctly identify and process them.
- Upload a product document with known updates. Then, query the updated content through conversation to confirm that the knowledge base has synchronized the latest information.
- For ambiguous queries or questions containing synonyms that users might ask, check the relevance of the system's recall results and the prompt's guiding capability.
Note: The values provided are common starting points. They 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.