Retrieval Failure in a Joint-Venture Document Repository

AI search may help with ranking and natural-language queries, but it cannot reliably infer authoritative revision status when the underlying records are contradictory. The repository currently lacks a stable document identifier, a controlled revision field, and a clear approval state. An AI tool trained on that mess would confidently return superseded or duplicate documents, increasing the dispute risk. Fix the identity and state model first. Assign a normalized document number and revision scheme; make revision and approval status mandatory metadata for controlled documents; distinguish controlled records from reference material; and mark superseded versions as superseded rather than deleting them. Design retrieval around facets-discipline, system, document type, revision, approval status, contractor-and configure search to prioritize approved current revisions. Preserve transmittal evidence because it supports the contractual dispute. Then consider AI-assisted classification or ranking as an enhancement, with human validation. The retrieval design should support two modes: precise lookup by document number and exploratory search by system or discipline. Without those controls, better search merely makes bad evidence easier to find.

Use a hybrid approach: a light controlled taxonomy for high-value faceting plus targeted free-text tags for local needs. The mandatory core for controlled documents should be small: unique document ID, title, discipline or system, document type, revision, approval status, originator, and retention class. Add project-specific fields only where they support a real retrieval or audit use case. For the 40% missing discipline metadata, do not rely on manual backfill alone. Run rule-based classification from folder path, document number patterns, vendor name, and title keywords, then sample-check accuracy. Prioritize backfill for active commissioning and contract-critical documents rather than the entire archive. Use a temporary "unclassified" state that is visible and reportable, not a silent default. Enforce mandatory fields at upload for new controlled documents and at revision approval; allow exceptions through a controlled exception process rather than turning fields into free text. This balances speed with trust. Overly rich taxonomy will be abandoned during the 11-week crunch, while purely free tagging will reproduce the current inconsistency.

Prioritize by contractual and safety exposure, not by folder completeness. Identify the disputed document families and any systems needed for mechanical completion, safety, or regulatory approval. For those, reconstruct the approval chain from transmittals, review comments, approval letters, and repository history. Create an evidence package per disputed family: document ID, revision, transmittal date, reviewer, approval status, and location of the authoritative file. Freeze those records with read-only permissions and a change log. In parallel, apply a quick normalization to active commissioning documents: standard revision labels, mandatory status, and superseded markers. Defer cleanup of low-risk reference manuals and closed vendor correspondence. Use a small cross-party team with document control, engineering, and contracts, and publish a daily exception list rather than a large migration plan. This approach is defensible because it preserves provenance and separates current approved records from historical or reference material. It also avoids the trap of deleting duplicates that may be evidence. The steering committee should see a risk-ranked burn-down of disputed and commissioning-critical documents, not a raw count of cleaned files.