What the documents look like
For product questions in Convenience fresh-food orders, this page examines FastGPT retrieval. Compare the material received with the sources used in the answer. The examples are hypothetical.
Constraints at the retrieval step
A new fresh-food order test says, “Individually packed sandwich, recorded for the day’s order batch.” This supports reading unit packaging and record scope, without stating units per transit carton. Inspect candidates for individually packed and the day’s batch, then compare modes with the complete sales-packaging question. This ordinary-text test provides no real storage or consumption advice, and a packaging label cannot be expanded into an unsupplied quantity specification.
Configuration and implementation checks
| Documented item or implementation check | Material and condition | Evidence action |
|---|---|---|
| Source scope (implementation check) | Separate unit packaging and handling batch | A carton record does not redefine the unit attribute. |
| Search modes: semantic, full-text, hybrid | Literal unit-pack wording and the packaging-form question | Inspect source roles returned by each mode. |
| Citation token limit | Day-specific batch and awaiting-review conditions | Retain timing scope and the unresolved state. |
| Minimum relevance | Inspect filtering under semantic search or reranking | A high-scored passage does not supply carton quantity. |
| Question optimization | Follow-up about units per carton | Keep product and batch without inventing a count. |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Associated knowledge bases | Knowledge bases associated with one search node must use the same embedding model. |
Search mode | Semantic, full-text and hybrid modes are available. |
Citation limit | Citation limits are measured in tokens. |
Minimum relevance | Range 0–1; active for semantic search or reranking. |
Question optimization | Resolves references and expands questions. |
Three common errors
A proposed near-expiry sheet says only “carton batch awaiting review.” Its level differs from individual packaging. An answer using carton wording to redefine the sale unit mixes batch handling with a product attribute; trace both sources. Inspect token-limited citations for omitted day-specific or awaiting-review conditions, avoiding conversion of a limited record into a confirmed state. The statements can coexist, without a retrieval ranking deciding which packaging level represents the actual product. A third error turns a batch awaiting review into a decided whole-carton disposal arrangement. Seek the actual handling decision for that batch. A pending check and an execution decision are different evidence objects; record granularity alone cannot establish the action.
How to verify the setup
The units-per-carton count and corresponding packing specification are absent. Identify that missing material rather than calculating a count from individually packed. For a carton-count follow-up, inspect whether optimization preserves the same product and batch without inventing a number. After obtaining the actual specification, repeat the question and trace added candidates and citations, recording sources separately from settings. Minimum relevance applies to semantic search or reranking; one highly scored result cannot replace packing details.
| Material context | Scope |
|---|---|
| Document types | Fresh Food Order Form, Spoilage Report, Near-Expiry Processing Sheet and Temperature Records |
| Candidate sources | POS Sales, Headquarters Ordering System, Store Inventory and Cold Chain Temperature Control Records |
| Technical baseline | FastGPT v4.17.2 |