This process leads from an empty installation to an initial response built on your
\ndocuments. Allow about ten minutes.<\/p>\n
On first launch, WivenLLM lets you choose the essentials:<\/p>\n
The language model provider.<\/strong> The engine is what writes the answers. The embeddings engine.<\/strong> It translates your documents into vectors. The engine Important :<\/strong> Changing the embeddings engine later makes the indexes The vector basis.<\/strong> The integrated local database is the default choice: nothing to Your account.<\/strong> Name and credentials of the first account, which is the administrator.<\/p>\n Your first workspace.<\/strong> A name is enough \u2014 for example, the name of a A workspace<\/strong> is the organizational unit of WivenLLM: it is he who Three ways:<\/p>\n The import process takes place in two stages, and this is the most common source of confusion. Until step 2 is completed, the document cannot be searched.<\/strong><\/p>\n \u2192 Documents<\/a><\/p>\n Write your question in natural language. The answer will appear as you type. Three tips that greatly improve the quality of the responses:<\/p>\n Open the space settings (gear icon next to its name). Three The mode of conversation.<\/strong> For strictly documentary use (client files, regulatory database), the mode The prompt system.<\/strong> The permanent instructions given to the model: its role, its The model.<\/strong> Each space can use a different model than the one \u2192 Conversations<\/a><\/p>\n
\nYou can choose a locally run model, an inference server from your
\norganization, or an external supplier. This choice then changes at any time.
\nmoment, including workspace by workspace.
\n\u2192 Choosing your models<\/a><\/p>\n
\nThe integrated version runs locally, requires no key, and is suitable for most...
\nuses.<\/p>\n\n
\nExisting spaces become unusable and require reindexing. This choice deserves attention.
\nto be done once, correctly.<\/p>\n<\/blockquote>\n
\nOnce installed, everything remains on disk. An enterprise database is justified on large systems.
\nvolumes or when it already exists in your home.<\/p>\n
\na client's or project's file.<\/p>\n
\n2. Understanding the main screen<\/h2>\n
\n
\n settings and conversations.<\/li>\n
\ndetermines which documents are available for consultation, which model meets the requirements, and according to which
\nInstructions: Two spaces share nothing unless you add the same document.
\nto both.<\/p>\n
\n3. Import a document<\/h2>\n
\n
\n file) and upload from there; ;<\/li>\n
\nfrequent:<\/p>\n\n
\n document library.<\/li>\n
\n sliced, vectorized and indexed.<\/li>\n<\/ol>\n
\n4. Ask a question<\/h2>\n
\ngeneration, followed by a mention Sources<\/strong> which unfolds the extracts used:
\nName of the document and exact passage.<\/p>\n\n
\n "Dupont?" is better than "cancellation deadline?".<\/li>\n
\n Indexing or search parameter problem.<\/li>\n
\n sent to the model; mixing the subjects degrades the responses.<\/li>\n<\/ul>\n
\n5. Adjust the workspace<\/h2>\n
\nadjustments produce most of the effect:<\/p>\n
\n\u2013 Cat<\/strong> \u2014 the template uses your documents And<\/em> his general knowledge.
\n\u2013 Request<\/strong> \u2014 the model responds uniquely<\/strong> from your documents, and
\n indicates that he does not know when they do not contain the answer.<\/p>\n
\nQuery is almost always the right choice.<\/p>\n
\ntone, what he should refuse, the expected format. This is the most effective lever
\nof the entire application.<\/p>\n
\nthe instance.<\/p>\n
\n6. Going further<\/h2>\n