Franchise filing management: document chunking for procurement

Review document chunking for procurement in Franchise filing management. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Franchise filing management works with franchise contracts, operations manuals, filing materials and information-disclosure documents. Its data sources include the commercial franchise information management system, enterprise licences, store lists and contract ledgers. Inventory procurement and supplier materials form one business-material group inside this category, and the common shapes are purchase lists, supplier directories, supply-term descriptions and attached licence documents. Responsibility for these materials is split across the franchisor headquarters, franchise legal staff and the person who liaises with the commerce authority. Updates are irregular, with some items refreshed annually. Because errors can lead to an order to file, franchise-contract disputes and administrative penalties, parsing and chunking must preserve the document identity and version that allow a chunk to be traced back to its source text, and must keep conditions, explanations and the content they govern inside the same unit where the structure allows it. Retrieval returns candidate content only and does not decide authorization; who may view, submit or execute remains a downstream business-process question.

Constraints at the document chunking step

Chunking follows the structure of the specific material. Purchase-list numbers, supplier names and quantity units are precise-query objects and suit full-text or hybrid retrieval; supply-term descriptions are natural-language content and suit semantic retrieval. One group of source data can map to several indexes, so a richer set of indexes helps recall, while a shorter index content sacrifices retrieval range. The reference limit is counted in tokens, and the minimum relevance value takes effect only under semantic retrieval or result re-ranking; neither acts as a permission filter over suppliers or stores. Hybrid retrieval merges multiple result paths through RRF, re-ranking requires a model and must be enabled by configuration, and image retrieval requires image input together with the corresponding model capability. These optional paths are not required to run on every request.

Hypothetical paper exercise. The source material to be checked reads: "This year, stores in the East China region are supplied with frozen dough by Supplier A, with a supply period from 1 January 2024 to 31 December 2024. The quotation is valid for 30 days." One draft sentence to be judged reads: "Supplier A can supply throughout the whole of 2024." Checking compares subject, period and field: the source gives a supply period ending 31 December 2024, and the draft's "whole of 2024" matches that period. The 30-day quotation validity is a different field and cannot be merged with the supply period. If the draft instead reads "Supplier A's quotation is valid throughout the whole of 2024," it conflicts with the 30-day quotation-validity field and is unsupported. Where a period is not stated, the item is recorded as pending verification.

Configuration and implementation checks

Configuration itemAssumed trial starting valueReason and comparison method
Search modeHybrid retrievalPurchase numbers need full-text hits and term descriptions need semantic recall; compare candidate differences for the same question under both modes
Reference limit3000 tokensTrial starting value only; compare against a lower value for candidate completeness, not treated as an official default
Minimum relevance0.3Trial starting value only; observe filtered candidates when semantic retrieval or re-ranking is enabled
Chunking methodSegment by clause and by list rowCompare whether a condition sentence and the content it governs stay in one chunk
Question optimizationEnabled for multi-turn follow-upsCompare consecutive-dialogue retrieval before and after enabling; can be disabled for short, explicit questions
Post-update reviewCheck training and index status after a changeUpload success does not mean retrieval has updated; compare the same query before and after the change

Three common errors

  1. Treating quotation validity as the supply period. Input the draft sentence "Supplier A's quotation is valid throughout the whole of 2024," compare it against the source's 30-day quotation-validity field, and observe whether the same field holds inconsistent values; a period not stated in the source is recorded only as pending verification.
  2. Treating a retrieval hit as authorization. Input one retrieval request, compare the returned candidates against the manually confirmed accessible scope, and observe whether recalled content is used directly for submission or execution; retrieval does not perform permission filtering.
  3. Treating upload success as an updated index. Input the same query after a change, compare candidates before and after, and observe whether old content is still recalled; actual training and index status must be checked separately.

How to verify the setup

  • Check one: take the source text "supply period from 1 January 2024 to 31 December 2024" and input the draft sentence "Supplier A can supply throughout the whole of 2024." Compare subject, period and field. If the period matches and quotation validity is not mixed in, the draft is judged supported; if the draft reads "quotation valid for the whole year," it conflicts with the 30-day quotation-validity field and is judged unsupported.
  • Check two: run semantic retrieval and full-text retrieval separately for "frozen dough" and "Supplier A," compare whether candidates cover the sample clause, and observe whether the number-type query produces a hit. If candidates are empty, record this as not recalled and do not infer that a rule necessarily returns empty.
  • Check three: after changing the same material, query again and compare candidates before and after, observing whether old content still appears. Upload success is not evidence that retrieval has updated; the actual state follows the check result.

Operational handover must keep clear which materials, operations and requests sit at which authorization state. Retrieval and forms are not authorization execution.