Fable
The Q&A Workshop arranged book finding, book filtering and book answering into an assembly line
North City's Q&A Workshop answered institutional, product, and contract questions for merchants. At first there was only one old clerk, flipping through materials from memory, searching while answering. As merchants grew in number, the old clerk began missing new documents. Similar questions were answered inconsistently, and citations often pointed to outdated versions. The workshop owner first bought a bigger bookcase and stuffed all the documents into it. Search got a little faster, but some questions needed keywords, some needed semantics, and some needed evidence reranking. He then let each clerk choose their own tools: some used the old index, some used the similarity cabinet, and some asked the agency directly. As a result, the pipeline was hard to trace and errors were hard to reproduce.
The new workshop master broke the work into an assembly line: documents enter storage, the retriever finds candidates, the reranker filters evidence, and the reader or generator organizes the answer.
Different customers can swap in different components. Customer service knowledge bases lean toward Q&A, legal materials emphasize citations, and product search requires hybrid retrieval and filtering.
The pipeline makes problems clearer. Whether the document never entered the database, retrieval missed it, or the generator made things up can finally be located segment by segment.
The workshop owner still has to manage index refresh, permission isolation, evaluation sets and latency budgets. Having all the components does not mean business answers are naturally reliable.
He understood that the whole Q&A apparatus is not a single mechanism call, but a composable, observable, replaceable knowledge-processing pipeline.
The new workshop owner also set up a small desk for verifying answers. Every time an assistant finished answering, the citations had to be placed back next to the original text for verification; if the evidence was too old, the permissions should not have allowed viewing, or the answer went beyond the materials, the entire pipeline had to be sent back for adjustment.
A few months later, the Q&A workshop landed a major client. The client did not want a talking assistant, but a process that could swap cabinets, swap rulers, check errors and prove provenance. Only then did the workshop owner realize that what was truly valuable was turning finding materials, screening evidence and organizing answers into governable operations. When the client asked to switch to a different set of bookcases or a different answerer, the pipeline could still hold, without tearing down the whole Q&A workshop. This allowed each renovation to start from the segment where the problem occurred.