What the documents look like
FBA preparation moves inventory ahead of demand into a platform warehouse, so inbound discrepancies and inventory age carry the main risk. Missed stock can reduce listing standing, storage fees can rise, and claims can fail. The materials for inventory procurement and supplier records therefore include purchase orders, supplier profiles, inbound plans, and compensation claims. They come from Amazon Seller Central, FBA inventory reports, freight forwarder tracking, and ERP purchase orders, and they update on a daily and weekly basis.
Document parsing and chunking works on how this same material behaves when it is split, indexed, and retrieved. Retrieval returns reference content only. A retrieved passage is not an authorization, an approval, or a signed agreement.
These materials commonly carry supplier names, purchase order numbers, model numbers, quantities, inbound status, and compensation terms. A conditional sentence and the object it constrains must stay in one segment. When they are split apart, the retrieved fragment loses its subject or its effective scope.
FastGPT organizes a knowledge base into library, collection, and data. The smallest search unit is the library, so a search runs across the whole library, while a collection only classifies data and does not affect search results. One set of source data can map to multiple vectors. During retrieval, vectors are recalled first, then the vector ID is used to find the source content in MongoDB, and entries that resolve to the same source data are merged with the highest score kept. Semantic search relies on vector distance. Full-text search suits numbers, model codes, and proper nouns. Hybrid search merges both paths through RRF. Rerank produces a score from 0 to 1. The minimum relevance value takes effect only under semantic search or with result reranking, and it is a filter threshold, not a permission or entity-isolation control. The reference limit is counted in tokens, not in items. Question optimization adds one model call before retrieval, which can improve multi-turn retrieval and adds latency.
Assume a paper exercise sample. The supplier record reads: "Supplier A submitted a quotation on March 1, 2026, the quotation is valid through March 31, 2026, and purchase order PO-2026-0310 remains unsigned." A draft sentence under review reads: "Supplier A has confirmed supply through March 31, 2026." The subject needed for judgment is Supplier A, the objects are the quotation and purchase order PO-2026-0310, the states are quotation valid and purchase order unsigned, and the effective scope is March 1 to March 31, 2026.
Constraints at the document chunking step
Chunking has to keep the states attached to the objects they describe. In the sample, the quotation validity date and the unsigned purchase order are two separate fields. A draft that converts quotation validity into a supply commitment draws on evidence the source does not contain. The correct comparison is between the draft sentence and the source fields for quotation and purchase order. When the source does not state a supply commitment, the draft sentence is recorded as unsupported, and the evidence gap stays open.
Retrieval does not carry authorization. A retrieved supplier record supplies reference content. Purchase order signature, supplier onboarding, and payment approval require downstream implementation. When the retrieval output carries no authorization field, the authorization state is recorded as pending verification, and a pending request is not an approval.
Upload success does not mean the index has updated. After material changes, the actual training state and index state need review. The comparison is between the upload record and the content returned by retrieval. When returned content still shows an older version, the index is recorded as not updated. Retrieval is not authorization, and forms collect values; business verification, approval, database writes, and external synchronization need their stated downstream implementation.
Configuration and implementation checks
| Configuration item | Setting or check on this page | Note |
|---|---|---|
| Search mode | Hybrid search | Purchase order numbers and model codes use the full-text path; natural-language questions use the semantic path; results merge through RRF |
| Minimum relevance | Hypothetical trial starting point 0.4 | A starting point only, compared against actual retrieved scores before adjustment; not an official default |
| Reference limit | Set in tokens | The documents use tokens as the reference limit unit and do not use top k |
| Question optimization | Decided per scenario | Enabled for multi-turn follow-ups and short questions; can stay off for already-clear questions |
| Chunk review | Condition and constrained object in one segment | Supplier name, purchase order number, state, and validity period are not split apart |
| Post-update review | Check actual training and index state | Upload success is not the same as a refreshed retrieval index |
RRF is a fusion step for hybrid results, not a switch to toggle. Image retrieval requires an image input and a model that supports it, and reranking requires a rerank model enabled under the settings. A library tied to a different vector model cannot be jointly associated, so the identifier is checked first, non-matching libraries are excluded, and actual observations are recorded.
Three common errors
- Treating a quotation validity period as a supply period. In the sample, the quotation is valid through March 31, 2026, and purchase order PO-2026-0310 is unsigned, while the draft sentence states confirmed supply. The check compares the draft sentence against the quotation and purchase order fields in the source. When the source states no supply commitment, the draft sentence is recorded as unsupported, and the gap stays open.
- Treating retrieval as authorization. A retrieved supplier record supplies reference content only, and purchase order signature, supplier onboarding, and payment approval need downstream implementation. The check compares the retrieval output against authorization fields. When the retrieval result carries no authorization state, it is recorded as pending verification.
- Treating upload success as an updated index. After material changes, the actual training state and index state need review. The check compares the upload record against the content returned by retrieval. When returned content still shows an older version, it is recorded as not updated.
How to verify the setup
- Input the sample source text and the draft sentence "Supplier A has confirmed supply through March 31, 2026." Compare the subject, objects, and states in the draft sentence against the quotation validity and unsigned purchase order fields in the source. The draft converts quotation validity into a supply confirmation and is judged unsupported; when the source states no supply period, the evidence gap is retained.
- Input purchase order number PO-2026-0310 into full-text search. Compare whether the returned fragment contains that number together with the unsigned state. When the returned fragment contains the number but the state has been split into another segment, chunking has not kept the condition and its constrained object together. When the return is empty, it is recorded as not recalled, and no failure cause is inferred.
- Input a clearly labelled paper empty result []. Check the downstream design for handling an empty reference and compare whether the downstream treats the empty result as no authorization or as a pass-through. When the downstream treats an empty result as an authorization to execute, it is recorded as a design defect. The empty result itself does not establish any business fact, and an empty result does not show that the search returns [] under every condition.
Each check states what is compared and what observable result satisfies the condition. No test, deployment, approval, refund, or other business action is claimed to have occurred.