Fine-tuning or RAG: what will improve product answers?
Suppose a sales assistant describes a product as available even though the order system now marks it unavailable. The team considers further model training. First, establish where the assistant got its information: did it use an old product sheet, or did it receive the current status and misrepresent it? That distinction determines what work is worth commissioning.
Syntalith
Syntalith proposes assessment and development of an AI application using product data. We start with answers employees correct and the information available to the model when it produced them. We then compare clearer instructions, a connection to a current source, and possible model customization. The buyer can inspect how answers change after each adjustment before agreeing on further work.
Is a current fact missing, or is the model using it incorrectly?
Further training on past answers will not provide today's order status. The application needs the source where the company actually maintains availability and a way to read it when preparing an answer.
The situation is different if the model received “unavailable” but told the customer the product could be ordered for immediate delivery. The information arrived, but the answer failed to preserve it. Review the task instructions and assess the existing model on similar questions. Recurring mistakes despite correct data may justify evaluating changes to the model's behavior.
What RAG provides, and what keeps it current
RAG is an approach in which an application retrieves relevant passages from company material and supplies them to the model along with the question. It can help when a salesperson asks about a feature described in a technical sheet, manual, or approved product description. The answer can then link to the document used.
Retrieval alone does not keep material current. If a new product sheet reaches a folder but the application still retrieves an old copy, the error remains. The project needs to establish how revised documents become available and how earlier versions stop being used for current answers. This also applies to saved results the application might display again.
For product availability, a direct lookup in the system maintaining that status may be more appropriate. There is no need to create a document and add it to a search index first. Depending on the question, the application can use technical descriptions or the current operational system. The team identifies the right source for each kind of information and decides what should happen if it cannot be retrieved.
When an employee's existing tool is enough
For occasional comparisons of approved product sheets, start with an existing tool. ChatGPT Work works with information, files, and connected apps available to the user. Claude Cowork supports tasks such as analysis and document creation using selected material and tools. Capabilities depend on settings and access.
Check whether employees can already prepare the answer they need in that environment. A separate application is worth considering when consistent data retrieval and source references need to become part of everyday sales work without someone selecting files for every case. The scope may involve connecting systems while continuing to use an existing model.
What fine-tuning changes
Fine-tuning means further training an existing model on reviewed examples of the expected behavior. It can help the model perform a defined task more consistently, such as preserving the distinction between confirmed availability and information that still needs checking. It is not a way to maintain current inventory status.
Before a trial, refine the instructions and assess whether the current model uses the supplied information as intended. If mistakes persist, the team identifies corrected answers and explains what the model should have done. Evaluation uses separate questions that were not used during customization, with the appropriate product information supplied.
A company may need both better access to data and more consistent behavior. Assess those parts separately.
Define the scope around the team's answers
When defining the project, your team identifies the sources for specifications and availability, along with the people responsible for updating them. Syntalith examines how information reaches an answer and compares the changes needed. The buyer can inspect a sample answer alongside the actual source from which the application obtained its information.
For an initial conversation about the application and model, describe an incorrect answer and identify where the correct information was available at the time. This lets us discuss the actual problem before choosing RAG or fine-tuning. We will agree together on how examples are used. See the Syntalith pricing page.
Match a model to the task you need it to perform
Describe where your current AI falls short. We will compare model customization options, data requirements and the cost of running the resulting system.
Private LLMs and fine-tuning