Knowledge Base Retrieval and Recall for Batch Record Audit

Data Characteristics for This Category

Batch record audit data primarily originates from pharmaceutical manufacturing process documents. These include batch production records, batch inspection records, deviation investigation reports, and change control documents. Data typically exists as scanned PDFs or structured data forms (e.g., Excel, XML). Update frequency aligns with batch production cycles, often daily or weekly. However, historical batch data is extensive and changes infrequently. Document structures are highly standardized, adhering to GMP guidelines. They contain numerous fixed fields such as batch number, product name, production date, expiration date, operator signature, equipment number, key process parameters (temperature, pressure, time), material batch, and inspection results. Field values are often numerical, date, boolean, or enumeration types, with clear and varied units (e.g., ℃, kPa, min, mg/L).

Constraints on Knowledge Base Retrieval and Recall

The standardized structure and explicit units in batch record data enable precise matching based on field names or units. This demands high accuracy in recall. Document update frequency dictates the knowledge base indexing strategy. The stability of historical data allows for less frequent full index rebuilds, while real-time updates for new batches require incremental indexing. The large volume of numerical and enumeration data requires the knowledge base to handle range queries, equality queries, and compliance checks against specific values and specifications. Free-text fields, such as operator signatures and deviation descriptions, necessitate semantic understanding capabilities. The low tolerance for errors in recall results requires precise adjustment of matching weights for key fields during similarity calculation to prevent "generalized" recall leading to misjudgments.

Configuration Guidelines

Configuration Item	Recommended Value	Rationale
`Chunk size` (Segment Length)	`300–500 characters` (characters)	In batch record documents, key information is often concentrated in specific paragraphs. Segments that are too long or too short hinder precise matching.
`Recall count` (Recall Count)	`Top 8–12 entries` (top 8–12 items)	Considering the complexity of batch records and potential related information, increasing the recall count improves coverage.
`Similarity threshold` (Similarity Threshold)	`0.78–0.85`	Batch record audits demand high accuracy. A high threshold helps filter out irrelevant recall results.
`Rerank result count` (Rerank Return Count)	`5 entries` (5 items)	Reranking initial recall results ensures the most relevant few results are displayed first.
`UPLOAD_FILE_MAX_SIZE`	`200 MB`	Scanned batch records can be large. Ensure the ability to upload large files.
`PARSE_FILE_TIMEOUT_SECONDS`	`600 seconds` (seconds)	Set sufficient timeout for parsing large PDF batch record files.

Common Pitfalls

Outputting a question in the chat window with the prompt "Knowledge base not selected" usually indicates that the knowledge base component is not correctly enabled or linked in the Agent configuration.
When no directly matching entries are found in the knowledge base, the model directly refuses to answer. This may be due to a Similarity threshold (Similarity Threshold) set too high or Recall count (Recall Count) set too low, preventing the recall of sufficient information for inference.
An empty result from the Google search plugin may relate to network configuration, an expired API Key, or incorrect plugin integration, limiting external search capabilities.

How to Verify Configuration

Upload a batch of typical batch record documents. Verify that file parsing succeeds without timeout errors under the configured UPLOAD_FILE_MAX_SIZE and PARSE_FILE_TIMEOUT_SECONDS.
Construct a series of precise and fuzzy queries targeting key fields and operational specifications within batch records. Observe the Recall count (Recall Count) and relevance of the recalled results. Adjust the Similarity threshold (Similarity Threshold) based on expert judgment.
Simulate audit scenarios. Input queries containing numerical ranges and unit information. Check if the recall results accurately hit relevant batch data. Verify the sorting priority of the most relevant information after Rerank result count (Rerank Return Count).

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.