What the documents look like
Parcel label and collection handoff work happens at the point where a parcel leaves the warehouse, and responsibility turns on carrier scan records. The core data comes from the courier system, WMS shipment records, GB/T courier service standards and carrier billing. Daily documents include electronic waybills, handoff manifests, collection scan records and exception item registers. These documents update per batch and per day.
For product material and selection questions, the material identities that matter are the shipment number, collection time, handoff batch and exception status. Chunking must keep a condition, its explanation and the content it constrains inside the same data group, and must preserve document identity and version so a retrieval result can point back to the source text.
Constraints at the document chunking step
FastGPT builds knowledge bases with an Embedding approach. The smallest search unit is the library; a collection is only a classification container and does not affect search results. One group of data can map to multiple vectors, and when several vectors resolve to the same original data, the vector score takes the highest value.
Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, hybrid retrieval merges multiple result paths through RRF, and Rerank performs a second ordering of candidate text. The minimum relevance value is a filter threshold that takes effect only under semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. The reference limit controls how many tokens are cited per search. Image retrieval requires a corresponding index to have been built at ingestion time. Rerank requires a model and must be enabled by configuration.
Libraries that use different vector models cannot be associated together. The identifier should be checked first and non-matching libraries excluded.
Hypothetical paper test sample
The following fictional material is used for this paper comparison. It does not involve a real customer or a real shipment number.
Source material one: "The handoff manifest number for this batch is JC-2025-0417, the handoff date is 17 April 2025, the carrier is Example Express, the handoff parcel quantity is 120 pieces, and the handoff status is handed over."
Source material two: "Collection scan records show that 118 pieces in the above batch completed collection scanning before 18:00 on 17 April 2025, and 2 pieces are registered as exception items, with the exception reason pending carrier confirmation."
Draft sentence to be judged: "All 120 parcels in this batch have completed collection scanning, with no exception items."
Comparison task: compare the draft sentence with the two source materials and decide whether its statements are supported. Source material two expressly records 2 pieces as registered exception items, while the draft sentence states there are no exception items. The two take mutually exclusive values on the same field for the same batch. The draft sentence also states that all parcels completed collection scanning, while source material two supports only 118 pieces as completed, with the remaining 2 pieces in an exception state pending confirmation. That part of the statement is not supported.
Configuration and implementation checks
The table below lists only settings documented in the official material that relate to document parsing and chunking on this page, or implementation checks to be performed. The values are hypothetical starting points for this paper comparison and are not official default values.
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | The handoff manifest contains the number JC-2025-0417 together with natural language description, so keyword and semantic recall both need coverage; compare hit results against semantic retrieval alone |
| Reference limit | 2000 tokens | The two source passages are short in total; this starting point is used to observe whether the handoff manifest and the exception item passage are both covered in full; compare citation completeness against 1000 tokens |
| Minimum relevance | 0.3 | This value takes effect only under semantic retrieval or result reranking and is used to filter low-relevance passages; compare the number of retained candidates for the draft sentence check against 0.5 |
| Rerank | Enabled | Performs second ordering of text when candidates are numerous; compare the ordering position of the relevant passage for the draft sentence against the disabled state |
| Question optimization | Disabled | The current check question has a clear subject, and disabling avoids one extra model call; compare retrieval latency and coreference resolution against the enabled state |
| Data and index check | Passage by passage | Confirm that both source passages have been trained and have generated indexes; check upload success and retrieval refresh as separate states |
Three common errors
- Treating upload success as retrieval already refreshed. If a source passage has been uploaded but training and indexing have not completed, retrieval may return an empty result or old content. Check action: inspect the data training state and the index field, compare the upload record with what is actually retrievable, and observe whether the retrieval result contains both source passages.
- Accepting the phrase "all completed collection scanning" in the draft sentence at face value. Source material two supports only 118 pieces as completed, with 2 pieces as exception items pending confirmation. Check action: extract the quantity and status fields from the draft sentence, compare them item by item with source material two, and observe whether the two agree on the same batch.
- Using minimum relevance to exclude unrelated business or unauthorized records. Minimum relevance is a filter threshold and does not perform permission or entity isolation. Check action: inspect the retrieval scope and authorization boundary, compare whether returned content belongs to the target library, and observe whether any cross-library or over-privileged return occurs.
How to verify the setup
- Input source material one and source material two, and compare the draft sentence "All 120 parcels in this batch have completed collection scanning, with no exception items" against the recorded text. Expected observation: the conclusion states that the draft sentence and source material two take mutually exclusive values on the exception item field, and that the claim of all scanning completed is not supported.
- Input the handoff manifest number JC-2025-0417 using hybrid retrieval, and compare whether the returned content contains source material one and source material two. Expected observation: both passages are recalled and the cited content can point back to the source text. If only one passage is returned, record this as incomplete recall and do not determine a cause of retrieval failure.
- Input an empty reference scope [] and check how the downstream design handles an empty result. Expected observation: the downstream explicitly records the empty result state and does not interpret the empty result as no exception items or as all scanning completed. This check does not claim that a fictional rule retrieval necessarily returns an empty array.
Libraries that use different vector models cannot be associated together. During the check, inspect the identifier first and exclude non-matching libraries, record the actual observation, and do not run cross-model retrieval to compare missed recall. Image retrieval and reranking are optional paths that require the corresponding input and model capability, and are not treated as executed on every request.