A model asked about an internal expense policy updated last month may fail because it has never seen the document. This is a knowledge gap: the information needed to answer is missing, whether it lives in a recent publication, a private wiki, or a data warehouse the model cannot query. section 4.5 section 4.5.1
Distinguish a knowledge gap from a behavior gap before choosing an adaptation technique. A knowledge gap concerns missing facts; a behavior gap means the model has the facts but gives the wrong format, tone, or level of detail. Fine-tuning for missing facts can waste compute without making those facts reliably usable. Retrieval alone can supply the right document while leaving an output-format problem unresolved. section 4.5.1 section 6.5
Table 4.4 Knowledge gap signals versus behavior gap signals
|
Knowledge gap
signals
|
Behavior gap
signals
|
|
Model says “I don’t
know” or hedges on a question that has a definite answer.
|
The answer is
factually correct, but in the wrong format (paragraph instead of JSON, prose
instead of bulleted list)
|
|
Wrong factual answer
with high confidence (a hallucinated date, name, or number)
|
The answer is
correct, but the tone is inconsistent across runs (formal one time, casual
the next)
|
|
References
documentation, APIs, or product names that do not exist
|
The model follows
the output schema sometimes and ignores it other times
|
|
Answers about events
after the training cutoff or about private internal documents
|
Succeeds on simple
cases and fails on complex cases that draw on the same underlying knowledge
|
|
Same query,
different model versions, different factual answers
|
Same query,
different prompt phrasings, large variability in style or structure
|
Compare signals of missing facts with signals of inconsistent presentation or behavior. section 4.5.1
Use a short diagnostic before building a retrieval pipeline:
- Take ten failing cases, paste the relevant source documents directly into their prompts, and rerun them. A jump in quality is evidence of a knowledge gap and a reason to build retrieval. section 4.4.6 section 4.5
- If quality stays flat, investigate behavior rather than assuming that more prompt iterations will solve the problem. For a format-only gap, try a JSON schema or constrained decoding before fine-tuning. section 4.4.6 section 4.5.1
- If retrieval helps, evaluate retrieval hit rate, answer faithfulness, and end-to-end accuracy separately. Each answers a different debugging question. section 4.10
Retrieval-augmented generation (RAG) supplies relevant documents at inference time, turning the answer into a reading-comprehension task. Fine-tuning instead targets behavior. Both can be useful in the same system; neither is a substitute for diagnosing the failure. section 4.5.1 section 6.5
Contract review
For a legal contract review system, retrieve current regulatory text rather than relying on fine-tuning to supply recent changes. Fine-tuning can separately teach structured output tables and the firm’s risk-rating taxonomy. This separation supports current information and consistent presentation. section 6.5
