Tagged “generation”
-
Answering in the reader's language
The question is in French and the corpus is in English. What that does to quotes, citations, defined terms and your faithfulness checks.
-
When the question contains a false premise
Asked why a policy allows fifteen days when it allows ten, a grounded model will explain the fifteen. How to make it check the premise first.
-
The model does not know what day it is
Documents talk about now and next quarter. The model has no clock. How to make time resolvable in a grounded answer instead of guessed.
-
When the answer is a procedure
A how-to answer someone will execute fails differently from a factual one. Dropped steps, merged variants, and lost conditionals.
-
Arithmetic the sources didn't do
A total the model computed appears in no document. Derived figures are new claims, they can't be cited, and they are wrong more often than they look.
-
What to log about a RAG answer
When an answer is wrong, you need to reconstruct how it was produced. The trace that makes that possible, and the privacy cost of keeping it.
-
Structured output from retrieved context
Extracting fields rather than prose changes the problem: schemas can be validated, and every field needs a way to say it was not found.
-
The shape of a good answer
Length, structure and ordering are design decisions with real effects — starting with putting the answer in the first sentence.
-
Decoding settings for grounded answers
Temperature and its neighbours do something different in RAG than in creative writing. Sensible values, and what they can't fix.
-
Streaming a grounded answer
Streaming makes a system feel fast and makes validation impossible, because you have published the answer before you could check it.
-
Summarising across many documents
Some questions need fifty documents, not five. The staged approach that handles them, and where accuracy leaks out of it.
-
When one retrieval pass isn't enough
Some questions can't be answered from a single search. How iterative retrieval works, and the loop control it needs to stay sane.
-
Multi-turn RAG
Follow-up questions break retrieval, because 'what about contractors?' contains almost nothing to search for. How to handle conversation.
-
Permissions at generation time
Retrieval can surface a document the asker is not allowed to read, and the answer will summarise it. Where access control has to live.
-
Prompt injection in retrieved documents
Retrieved content goes into your prompt as instructions the model might follow. Why RAG makes this worse, and what actually reduces the risk.
-
Quoting versus paraphrasing
Verbatim quotes are checkable and stiff; paraphrase reads well and can drift. When each is right, and how to combine them.
-
When sources disagree
Two retrieved documents give different answers. What a good system does about it, and why silence is the worst option.
-
Checking citations after generation
A validation pass that confirms every claim is supported before the answer ships. What it can catch cheaply, and what it can't.
-
Citations that point somewhere real
Asking a model to cite its sources is easy. Getting citations that actually support the claims attached to them is a separate problem.
-
Teaching a model to say I don't know
Abstention is a feature you build, not a default you get. Where the decision belongs and how to keep it from swallowing good questions.
-
Grounding instructions that work
The few sentences telling the model what to do with retrieved material have an outsized effect. What belongs in them, in what order.
-
Why the model ignores your context
You retrieved the right documents and the answer came from the model's training instead. The causes, and what actually fixes each one.
-
How many chunks to include
More retrieved context is not reliably better. Why adding chunks stops helping, and how to find where that happens for your system.
-
Assembling the prompt
Retrieved chunks, the question, the instructions and the history all go in one prompt. How you arrange them changes the answer.
-
What happens after retrieval
Most RAG writing stops once the right documents are found. That is the halfway point — here is the rest of the pipeline, stage by stage.