What the documents look like
Violation appeals deal with penalties that have already been issued. The materials cluster around four document types: the violation notice, the livestream recording, product qualification files, and the appeal explanation. Each violation generates its own batch of files on an irregular schedule, so parsing and chunking must preserve document identity and version, allowing a later conclusion to be traced back to a specific notice or a specific qualification file.
The primary task for product materials and selection questions is to organize product qualifications, the product references inside violation notices, and the product descriptions inside appeal explanations into retrievable chunks for downstream question answering. A violation notice typically binds the violation facts, the penalty basis, the appeal deadline, and the product object into one set of constraints. Splitting these apart leaves a fragment that cannot support a judgment on its own.
Product qualification files commonly carry numbers, model designations, and validity periods. Semantic retrieval is unstable on exact strings of this kind, while full-text retrieval suits numbers and proper nouns. One group of original data can map to multiple indexes: vector original data is stored in the MongoDB dataset.datas table, and an indexes field records the array of vector IDs it corresponds to. During retrieval, vectors are recalled first, then the original data is fetched from MongoDB by vector ID; when the same group of original data is matched, the results are merged and the vector score takes the highest value. Whether an image can be retrieved depends on whether an image description index or image vector index was generated at ingestion time, together with model capability.
Constraints at the document chunking step
Chunking follows the structure of each material. The violation facts, penalty basis, appeal deadline, and product object in a notice belong to one constraint group. Qualification validity periods and product numbers belong to another. A chunk that separates a deadline from the notice it belongs to cannot answer a question about that deadline.
Document identity and version must survive chunking. When a qualification file is replaced, the retrieval result must be traceable to the new version rather than the old one. Upload success and index refresh are separate facts and are recorded separately.
The retrieval mode affects which questions can be answered. Number and model queries are better served by full-text or hybrid retrieval. Natural-language questions about product selection are better served by semantic retrieval, with the caveat that semantic similarity does not perform entity isolation or permission filtering. A similarity threshold is a relevance filter, not a category filter, and it does not exclude unrelated products or unauthorized records. Retrieval is not authorization.
Image retrieval is not a separate capability outside the knowledge base. It adds an image input path to the existing search chain, and it requires either a configured vision model or a vector model that accepts image input. The optional paths in the overall retrieval flow are not all executed on every request.
Hypothetical paper exercise sample: a material labeled "Violation Notice for a Certain Store, March 2025" contains the text "This notice concerns product number A-1024; the violation fact is that the title contained a prohibited term; the appeal deadline is seven days after delivery of the notice." In the same batch, a product qualification file contains the text "Number A-1024, qualification valid until December 2025." The draft sentence to be assessed is "This product qualification had already expired at the time of the violation." Comparison against the two source texts shows the qualification validity period covers the month of the violation, so the expiry status asserted by the draft sentence is not supported by the source text and remains pending verification. A field proving the validity period, or a separate explanation, would be needed.
Configuration and implementation checks
| Check item | Setting or action relevant to this page | Observation and comparison |
|---|---|---|
| Document identity and version | Retain notice number, product number, qualification version | Whether a chunk can be traced back to specific source text |
| Segment completeness | Split according to notice and qualification structure | Whether violation facts, deadline, and product object stay in the same segment |
| Index creation | Confirm whether text, image description, or image vector indexes were generated at ingestion | Compare against the indexes record in dataset.datas |
| Retrieval mode | Use full-text or hybrid retrieval for number-type questions | Hit situation for numbers and model designations |
| Reference limit | For this hypothetical trial, set a starting value covering the tokens of one notice plus one qualification file, on the reasoning that this combination forms the material needed for one complete judgment; compare against a smaller value covering only a single notice | Whether the reference covers both materials |
| Post-update review | After a material change, check actual training and index status | Record upload success and retrieval refresh separately |
Three common errors
- Treating upload success as index refresh. The input is a new version of a qualification file. Compare the upload record with the chunk content returned by retrieval. If retrieval still returns the old version text, the index has not reflected the change, and training and index status need to be rechecked.
- Treating semantic similarity as permission or entity filtering. The input is a number query for one product. Compare whether the returned content contains other product numbers. If unrelated product chunks appear, the similarity setting does not carry entity isolation, and the retrieval scope or manual review layer must handle this separately.
- Treating a pending appeal request as authorized. The input is an appeal explanation labeled "pending review." Compare its status field with an already effective notice. If the draft sentence claims on this basis that the appeal has been accepted, the claim is unsupported and must remain pending verification.
How to verify the setup
- Input the product number A-1024 from the hypothetical sample together with the violation notice source text. Compare whether the returned chunk contains the violation facts, the deadline, and the product object at the same time. The segment completeness check is satisfied only when all three appear in the same segment and can be traced back to the source text.
- Input the draft sentence "This product qualification had already expired at the time of the violation" from the hypothetical sample. Compare it against the validity period field in the qualification source text. If the source validity period covers the month of the violation, the draft sentence is unsupported. If the source text lacks a validity period field, record this as an evidence gap.
- Input an empty reference list []. Check the downstream design for handling empty results and confirm that empty results have a defined handling path. Do not infer from this that a fictional rule retrieval necessarily returns empty. For libraries across different vector models, check the identifiers and exclude those that do not match, recording the actual observation. Do not perform cross-model retrieval to compare missed recalls.