Document search must compare the user's query with the documents, then return the appropriate accessible revision. Comparing documents with the first document in a collection is not query search. Adding a language model does not correct that retrieval mistake.
Establish authority before ranking
For a hypothetical construction project, attach a document ID, project, revision, approval status and access rules. An older drawing can match the query closely and still be the wrong operational reference. Show revision and approval status beside a result, and identify superseded material clearly.
A small, runnable lexical baseline
This Python example uses only the standard library. It ranks approved documents by shared words with the query. It is a demonstration of query handling, not semantic AI search or a production document-management system.
import re
def words(text):
return set(re.findall(r"[a-z]+", text.lower()))
documents = [
{"id": 1, "approved": True, "text": "Approved structural drawing"},
{"id": 2, "approved": True, "text": "Site safety procedure"},
{"id": 3, "approved": False, "text": "Draft structural drawing"},
]
query_words = words("structural drawing")
ranked = sorted(
((len(query_words & words(d["text"])), d["id"])
for d in documents if d["approved"]),
key=lambda item: (-item[0], item[1]),
)
results = [doc_id for score, doc_id in ranked if score > 0]
print(results) # [1]It omits stemming, OCR, synonyms and real permissions. A vector implementation must transform the query in the same fitted feature space as the documents; scikit-learn text feature extraction explains text feature extraction. Document-to-document similarity alone does not process the user's query.
Protect the retrieval boundary
Filter by project and permission before returning content. OWASP authorization guidance recommends checking access on every request. A hidden result can still leak through a generated answer if unauthorized text enters the model context.
How should acceptance work?
Test exact references, vague queries, no matching document, contradictory revisions and users from different projects. Require source and revision in answers; allow the system to decline an unsupported answer. Scope AI integration around these tests rather than an invented accuracy percentage.