Franchise fee settlement: product questions with hybrid retrieval

Review hybrid retrieval for product questions in Franchise fee settlement. Check sources, input boundaries, error handling and repeatable test evidence.

What the documents look like

For product questions in Franchise fee settlement, this page examines FastGPT hybrid retrieval. Compare the material received with the sources used in the answer. The examples are hypothetical.

Constraints at the hybrid retrieval step

The franchise-fee example’s invoice attachment states that an invoice was issued. That written status is the reading target, without establishing actual payment or receipt of funds. Mark the original status wording, check candidates against the correct fee and rerank using the complete question. Invoices, ledgers and payment plans serve different purposes; a nearby settlement passage need not support invoicing. Preserve source identity and the scope of what the passage actually supports. Retain separate sources for invoicing wording and payment evidence, so a final summary does not compress “invoice issued” into “funds received.” Reranking checks retrieval and order; the action verb used in the answer still needs verification against the corresponding supporting record.

Configuration and implementation checks

Item or implementation checkInput or applicable conditionTraceable check
Fused candidatesInvoice-issued wording in the attachmentKeep it separate from payment or receipt evidence.
Complete questionParty A in the invoice versus party B in the payment planCheck identity and correspondence to the same fee.
Minimum relevanceInspect filtering under semantic or reranking conditionsA high score cannot rewrite the payer in the original.
Citation token limitInvoice status and source identityDo not compress invoice issuance into received funds.
Missing-source retestNo receiving account in the available invoiceObtain an authorized source without inferring an account from party names.

Documented conditions

Documented item / implementation checkCondition
Hybrid retrievalVector and full-text results use RRF fusion.
RerankingReranking reorders candidates using the full question.
Minimum relevanceFiltering applies to semantic search or reranking.
Citation limitCitation content is limited by tokens.
Question optimizationCompare subjects and restrictions after optimization.

Three common errors

Party A and party B are named as payer in the proposed invoice and contract payment plan respectively. Retain both identities. Inspect candidate completeness before checking whether the documents concern the same fee and applicable versions. A longer or higher-ranked contract passage cannot rewrite the invoice party. The difference can remain a review item; establishing the actual payer needs corresponding business evidence, beyond a reranking decision.

How to verify the setup

The available invoice attachment supplies no receiving-account information. The model should not fill it from party names or the title of a deposit ledger. Inspect candidates and final wording for an account borrowed from an unrelated source. Once the corresponding authorized material is obtained, repeat the question and trace the new citation. Candidate order does not establish an account’s authenticity, ownership or completed settlement; preserve those distinctions in the review record.

Material contextScope
Document typesFranchise Fee Invoice, Deposit Ledger, Deduction Notice and Contract Payment Plan
Candidate sourcesFinancial System, Franchise Contract, Bank Statement and Store Violation Records
Technical baselineFastGPT v4.17.2