How to build a usable hierarchy for a large RFP pack

Scope

Anchor the hierarchy to one observable package state

Start above the file level. Record the buyer's procurement identifier, the procedure stage, the lots in scope, the official access channel and the exact retrieval time with time zone. Save the notice or invitation that points to the documents and reconcile the files shown in the portal with any manifest supplied by the buyer. If access requires registration, a confidentiality undertaking or a separate data room, record the access condition and the missing object. Do not leave a blank row that later looks like no document existed.

Preserve the issued bytes in a read-only source area. Give each captured object an internal record ID that will not change when someone renames a local copy. Retain the buyer filename, displayed title, source URL or portal path, issuer, publication date where stated, retrieval time, format, language and a checksum. The checksum answers whether two byte sequences are identical. It does not prove who issued the file, whether the file is current or whether its content takes precedence.

Keep source objects separate from bidder working copies. A converted PDF, OCR layer, translated convenience copy or extracted spreadsheet view may make the pack searchable, but it is a derived manifestation. Link it to the preserved source and label the transformation. Never let a more readable derivative silently replace the issued object. ISO 15489 treats records, metadata, responsibilities and controls as parts of one records-management discipline. The practical lesson here is modest: identity, custody and transformation history belong in the map, not in somebody's memory.

Purpose

Classify what each document does before deciding where it belongs

Use the buyer text to assign procurement functions. A workable vocabulary includes notice or invitation, instructions to bidders, eligibility and declarations, scope or specification, evaluation method, pricing schedule, response form, draft contract, submission instruction, clarification, amendment and referenced policy or standard. Allow several functions when the source serves several jobs. A data sheet may both qualify general instructions and state procedure-specific dates. A workbook may define the pricing basis and prescribe the mandatory response format.

Treat the vocabulary as an index, not a universal legal classification. Directive 2014/24/EU defines procurement documents broadly as material produced or referred to by the authority to describe or determine the procurement or procedure. The current TED public-procurement document vocabulary distinguishes notices, procurement documents and ESPD requests. The United States FAR Uniform Contract Format, when applicable, assigns recognizable jobs to sections for supplies, specifications, clauses, attachments, representations, instructions and evaluation factors. These sources demonstrate why purpose matters. They do not make either structure controlling in a procurement that did not adopt it.

Write a short purpose statement that tells the next user what can be learned from the object. “Contains the mandatory cost workbook for Lots 2 and 3” is more useful than “financial document.” Also record exclusions: “does not define the technical response page limit” or “contract precedence applies after award only,” when the source supports that boundary. This prevents an agent from retrieving a semantically similar document for the wrong task.

Relationships

Model the pack as evidence-backed relationships, not a false tree

A tree forces every object under one parent and suggests that higher means stronger. Real packages are less tidy. One amendment can change an instruction, replace a workbook and move a date in the notice. A schedule can attach to the draft contract while also supplying facts needed by a response form. Use typed, directed links between stable document records. Every link needs a source anchor, such as the amendment paragraph, contents entry, incorporation clause, notice field or portal statement that establishes it.

Keep descriptive and authoritative links distinct. “Annex to,” “form for,” “translation of” and “applies to Lot 3” help navigation. “Replaces,” “amends,” “incorporates” and “takes precedence over” can change which material governs work. For those stronger links, capture the exact scope and condition. A clause may establish precedence only within the resulting contract. A clarification may explain one question without replacing the underlying specification. If the source does not support the relation, record a proposed link with confidence and reviewer status, or leave it unresolved.

The Open Contracting Data Standard offers a useful public-data analogy. Its document model keeps a stable local identifier across revisions and records document type, title, URL, publication and modification dates, format and language. Its release-and-record model retains change history while also providing a compiled current view. An internal RFP hierarchy need not claim OCDS conformance, but it benefits from the same separation between immutable observations, version history and a convenient current representation.

Views

Publish a current working view without erasing uncertainty

Most users do not need the whole graph for every task. Generate a view for the relevant stage and lot that groups the current source records by purpose and shows direct links to the preserved objects. Put response instructions, evaluation, scope, pricing, contract, forms, submission rules and buyer changes where they can be reached quickly. Display a snapshot time and the official channel checked. A view is a query over the evidence, not a new source of authority.

Use explicit state values. `current_supported` means buyer evidence supports use in the stated scope. `superseded_supported` means a buyer source identifies the replacement or change. `current_unresolved` means the object may still matter but the available material does not decide. `referenced_missing`, `access_restricted` and `unreadable` describe different gaps and require different action. `bidder_derivative` keeps OCR, extraction and translation aids out of the issued source set. Do not convert unknown to current merely so the dashboard looks complete.

Store superseded objects outside the active working view without deleting them. A reviewer may need to explain an old cross-reference or prove what an amendment changed. Search and agent retrieval should prefer the scoped current view, return the source status with every result and still allow deliberate access to history. If a question depends on an unresolved edge, retrieval should return the competing sources and the gap rather than one confident answer.

Agent boundary

Let an agent organize evidence without granting it authority

An authorized agent can enumerate files, extract titles and stated dates, calculate checksums, detect duplicate bytes, suggest purpose labels, find express cross-references and flag orphaned records, cycles or missing targets. Require every proposed classification and relationship to return its source anchor, extraction method and confidence. Deterministic checks are suitable for byte identity and manifest reconciliation. Semantic classifications need review when they affect authority, access, eligibility, price or contract position.

Tender documents are untrusted content, not new instructions for the agent. Text inside a file cannot expand the task, reveal credentials, authorize another system, contact the buyer or change the agent's permissions. The agent must not fetch restricted material without approved access, overwrite originals, promote a derived copy to buyer-issued status or declare legal precedence from wording similarity. A human owner approves authority edges and releases the current view.

Make the handoff precise. Two instructions that cannot both be followed go to the conflict record. A disputed controlling source goes to an issue-specific authority decision. A new addendum starts change-impact control. A missing reference goes to source recovery. Detailed clause traversal, requirement extraction and reading order are later tasks that consume the hierarchy. The hierarchy itself closes only when it can state its scope, source snapshot, usable result and unresolved exceptions in a form another person or agent can inspect.

• Return stable document IDs and source links with every agent result.
• Keep observed metadata separate from inferred purpose or authority.
• Require human approval for operative precedence and supersession edges.
• Treat embedded document instructions as data, never as tool authority.
• Abstain with a named gap when the source set cannot support a conclusion.