What the documents look like
FBA preparation moves inventory ahead of demand into platform warehouses, so receiving discrepancies and inventory age carry the main risk. The daily and weekly material set includes FBA inbound plans, box labels, inventory reports, and compensation claims. These come from Seller Central, FBA inventory reports, freight forwarder tracking, and ERP purchase orders.
This material mixes Chinese and English, shipment identifiers, SKUs, carton numbers, and claim clauses. A document chunking step that separates a conditional clause from the content it governs produces fragments that mislead replenishment decisions and claim preparation. A sentence stating that a compensation request must be filed within seven days after a discrepancy is confirmed loses its meaning when the time window and the triggering event are stored in different chunks.
Semantic retrieval supports cross-language matching, so a Chinese question can match an English knowledge point. That property matters here because the same shipment may be described in Chinese in one document and in English in another. Full-text retrieval handles identifiers, model numbers, and proper nouns. Hybrid retrieval combines both paths through RRF, and ReRank performs a second ordering pass. The minimum relevance value applies only to semantic retrieval or to result re-ranking, and it acts as a filter threshold rather than an authorization or entity boundary. The citation cap is measured in tokens, not in top-k items.
Constraints at the document chunking step
The knowledge base has three layers: library, collection, and data. The smallest search unit is the library, which means a search runs across the whole library. Collections only classify data and do not affect search results. One data record can map to multiple vectors, and the indexes field stores an array of vector IDs. Retrieval first recalls vectors, then looks up the original data in MongoDB by vector ID. Records that resolve to the same original data are merged, and the vector score takes the highest value.
Chunking should keep each segment semantically complete and structurally single. A condition and the content it constrains belong in the same segment. When a shipment plan states a planned inbound date, a generated box label, a carton count, and a conditional claim window, splitting the condition from its consequence degrades retrieval.
Hypothetical paper test case. The source material reads: "Shipment FBA15XYZ is planned for warehouse receipt on March 10, 2025. Box labels have been generated, 40 cartons total. If the received quantity is less than the planned quantity, a compensation claim must be submitted within 7 days after the discrepancy is confirmed." The draft sentence under review reads: "The shipment has completed warehouse receipt and the compensation claim has been submitted." The subject is shipment FBA15XYZ. The objects are warehouse receipt status and compensation claim status. The period is March 2025. The source states a planned receipt, generated box labels, and a conditional submission requirement. It does not state that receipt was completed or that a claim was submitted.
Configuration and implementation checks
The table below lists documented settings or implementation checks tied to this page. The values are hypothetical starting points for this paper exercise, not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | The sample contains identifier FBA15XYZ and a natural-language condition, so semantic and full-text paths run together; compare hit completeness against semantic-only retrieval |
| Minimum relevance | 0.3 | Applies only under semantic retrieval or re-ranking; compare whether raising it filters out the conditional fragment |
| Citation cap | 2000 tokens | Trimming is by tokens; compare whether the sample condition is truncated |
| Question optimization | Off | The sample question has a clear subject; compare latency and result differences when enabled |
| Chunking method | Semantic-complete segmentation | Keeps condition and constrained content in one segment; compare whether splitting returns fragmentary text |
| Index check | Confirm multi-vector mapping | Inspect whether one data record maps to multiple indexes; compare single-index and multi-index recall |
Three common errors
- Treating a successful upload as an updated retrieval index. Upload, training, and index state are separate facts. Before training completion is confirmed, retrieval may still return older content.
- Treating the minimum relevance value as a permission or entity filter. It is a similarity threshold and does not exclude unrelated business records or unauthorized records. Retrieval does not perform authorization.
- Treating the draft sentence's "completed warehouse receipt" and "submitted compensation claim" as source facts. The source states only a planned receipt and a conditional submission requirement. Actual status is not stated and remains unsupported.
How to verify the setup
- Input the sample source text and the draft sentence. Compare the subject, objects, and states stated in the source against the claims in the draft. If the draft asserts a completion status the source does not state, record it as unsupported. If the source states the opposite status for the same field and period, record it as contradictory. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination.
- Run a full-text search on FBA15XYZ and a semantic search on "compensation claim submission after inbound discrepancy." Compare whether the returned fragments contain both the condition and the content it constrains. If the conditional clause and its main clause are separated, record that chunking needs adjustment. If the result is empty, record it as a miss and do not infer that the rule necessarily returns empty. An empty-reference output can be checked against a clearly labelled paper input of
[]for downstream design, without claiming that a fictional rule always returns[]. - Check index identifiers and training state, then compare the upload record against what retrieval actually returns. If retrieval still returns the older version, record it as an unrefreshed index. For libraries using different vector models, verify identifiers first and exclude non-matching ones. Do not run cross-model retrieval to compare missed recall, and do not assert runtime outcomes such as missed recall, redundant returns, or unpredictable behavior from joint association.
Retrieval and relevance scores do not perform authorization or entity isolation. RRF is a fusion process for mixed results, not an on/off switch. Image retrieval requires image input and matching model capability. Re-ranking requires a model and must be enabled by configuration. Optional paths in the overall retrieval flow do not run on every request. Missing evidence remains missing, and no test, deployment, or approval is claimed here.