How to record and fix a wrong knowledge-base answer
When a knowledge-base answer is wrong, the owner records the question, the wrong reply, and the correct source on one line, then decides whether to correct the current document, recut that passage, or retest retrieval. Citations decide whether a reply may leave draft with a path, a version, and a section. Retrieval evaluation asks whether a prepared question returned the passage. This line is what happens after someone marks a reply wrong.
On this page
- Keep the intake apart from citations and retrieval tests
- Name who may mark a reply wrong
- Put the question, the wrong reply, and the correct source on one line
- Correct the current document when the source sentence is wrong
- Recut, or fix the prompt, when the context was used and the conclusion was wrong
- Retest retrieval when the passage did not come back
- Have the owner write the document when no source exists
- Mark the reply fixed only after the route is done
Keep the intake apart from citations and retrieval tests
Citations ask whether a reply, before it is approved, carries a storage path, a version mark, and a section, and whether the owner opened that section and found the sentence there. Retrieval evaluation asks whether a prepared question returned the current passage, and it records a hit or a miss. The intake asks something else: after a reply is marked wrong, which line holds the question, the wrong reply, and the correct source, and whether the next step is to change the document, change the chunk boundary, or retest retrieval. The three stay separate. See How to attach sources to a knowledge-base reply and How to evaluate knowledge-base retrieval.
Which file counts as current, and who may mark an older draft as withdrawn, stays in the versioning note. Chunk boundaries stay where they are set, before indexing. See How to manage document versions in a knowledge base and How to choose chunk boundaries for a knowledge base. Permissions stay with the groups the file store already uses. See How to set permissions for an enterprise AI assistant. The delivery scope sits under Internal knowledge base and assistant.
Name who may mark a reply wrong
The person who marks a reply wrong is named. The name is the owner of this set of documents, or a reviewer the owner has already named on the same line. The reviewer must be able to open the path the wrong reply cites. If they cannot open it, someone who can open it marks the line, and the line says why the person changed. The date is written when the mark is made. A sentence in the group chat that says "this is wrong" is copied onto the line first. Without the line, the intake has not started.
The same reply may still be a draft, or it may already have been sent to a colleague. Both use the same line. If it was sent, the line records who received it and which channel carried it, so a corrected sentence can be sent again later. If it was not sent, the line says "not sent". Both states sit in fields, so the date of a later send can be matched to them.
Put the question, the wrong reply, and the correct source on one line
The line always holds the question the colleague asked, the wrong reply as it was worded at the time, and, when a correct source exists, the storage path, the version mark, and the section, plus who marked it and on which day. The wrong reply is copied as it stood. The three source fields match the citation note, and opening them shows that section on the store. See How to attach sources to a knowledge-base reply.
Read on 26 October 2026, Microsoft Learn: Large language model end-to-end evaluation describes groundedness, sometimes called faithfulness, as whether the response is based completely on the given context and does not use information outside that context. The page describes correctness as how far the response is factually accurate. The page shows no publication date. It is written for evaluating retrieval-augmented generation on Azure. A knowledge base can use the same distinction to choose a route. The page does not require a numeric score before a reply can be marked wrong.
Correct the current document when the source sentence is wrong
The page says that when correctness is low, first check whether the passages given to the model are factually correct and whether the source data is biased, and correct the source documents or content when that is needed. The intake therefore has one route: if the correct source, once opened, disagrees with a fact the business has already checked, the route is written as "correct the current document". Which file changes, who marks the old draft as withdrawn, and when the old chunks leave the index, follow the versioning note. See How to manage document versions in a knowledge base.
The line can be marked fixed only after the current document has been changed and marked current again. A line that only says "we will change it later" still counts as unfixed. If the reply was already sent, the new sentence is sent again, and the same line records the date of that send. The new sentence still has to open on a path, a version, and a section, under the same citation approval. The previous reply being wrong does not skip that approval.
Recut, or fix the prompt, when the context was used and the conclusion was wrong
The page says that high groundedness with low correctness means the model used the context and still gave an incorrect response. The cause can be a misuse of the context, or a problem in the source data. The page's illustration is a response that says Einstein developed quantum mechanics, when the context only mentions Einstein and quantum mechanics separately. The page calls that response grounded and factually incorrect. That illustration is the page's own example.
When the intake meets that pattern, it first looks at whether the passage the wrong reply used was cut away from the condition that governs it. If the condition was cut away, the route is "recut this passage". How the boundary is chosen is in How to choose chunk boundaries for a knowledge base. If the passage is still whole and the sentences are correct, and the reply joined two sentences into the kind of wrong conclusion the page describes, the route is "change the prompt the reply relied on", with the path of the prompt and the date. The page lists adjusting the prompt as something to evaluate. This line records the route. If a score is recalculated, that stays on the evaluation page's own method. This line does not gain a new score.
Retest retrieval when the passage did not come back
If the question has a passage in a known source, and the wrong reply did not use that passage, or it returned a different file, the route is "retest retrieval". The retest uses the questions and the current text already prepared in the retrieval note, and it checks whether this run returned that passage. See How to evaluate knowledge-base retrieval. This line does not invent a hit rate, and it does not turn one wrong answer into a score for the whole library.
The page says low groundedness means the model did not treat those passages as relevant, and that the next evaluation can be whether to add data, change the chunking strategy or the chunk size, or adjust the prompt. If the intake sees both "not returned" and "the passage was cut apart", both routes are written, and the line says which one is first. The first one is finished before the next one starts. The page also says that when correctness is low, inaccuracies in the model itself can be addressed with more factual context or with fine-tuning. Fine-tuning stays among the page's evaluation options. The route on this line is written first as the document, the chunk, the prompt, or retrieval. Until a separate sheet says where training data would come from and who would approve it, this line does not say fine-tuning.
Have the owner write the document when no source exists
If the three source fields are blank, the status is "no source". The owner writes the document, or names who will write it, places it in a folder already in scope, and marks it current. Only then does this line gain a path, a version, and a section. How the folders were chosen is in How to build an enterprise knowledge base. Until the file is marked current, the line cannot be marked fixed. The assistant does not cite a corrected sentence from a draft as the rule.
If the colleague's question was never inside this set of documents, the line says "out of scope" and names who should answer it. A question outside the scope stays with that owner. It is not rewritten into the knowledge base in order to produce a reply. "No source" means the scope includes the topic and the document is not there yet. "Out of scope" means these folders were not meant to answer the question. The two statuses stay in separate fields so a later check can tell them apart.
Mark the reply fixed only after the route is done
If the question, the wording of the wrong reply, the correct source or "no source", or the route (correct the current document, recut, change the prompt, retest retrieval, or write the document first) is blank, the line stays at "awaiting intake". A promise in the group chat still has to land on this line before the item has a place.
Before the line is marked fixed, three things are true: the line has the question, the wrong reply, and a source or "no source"; the route names the document, the chunk, the prompt, or retrieval; and the owner has signed after that route was finished. When those three are there, the line is marked fixed. A reply that was already sent has the date of the later send on the same line. Without that date, it stays unfixed.