Skip to main content
Every Digital Twin has a knowledge layer: the files people upload, indexed so the right passages can be found and handed to the AI when it answers. This page explains each setting that shapes it. You find these settings in two places:
  • Instance Settings › Knowledge — sections Personal Files and Memory, Ingestion and Retrieval. Display RAG/KAG Search Details is under Instance Settings › Conversation › Display Details.
  • The admin editor’s RAG and Files tab (Admin › Instances › Edit), which also holds Chunk Overlap (%) and the upload rules.
For what people see in chat, see Knowledge Modes.

​
How Pria searches files

Each person chooses how Pria uses files when they ask a question. There is no Digital Twin-wide default mode: the choice belongs to the person.
Runs the ordinary search and a search over the knowledge graph, then merges the two result lists. It finds connections the ordinary search alone misses (“Who reports to whom?”, “Which course covers which topic?”). When the vault in use has a knowledge graph, every question uses it unless the person chose RAG.
Each question searches the files in scope by meaning, and the best-matching passages go to the AI with the question. This is the only search mode where no knowledge graph is available.
A switch that works with either mode above. Pria returns the matching passages as they are, with citations, and no AI-written answer.
No file search. The AI answers from its training and the conversation alone.
People change their mode in the chat’s Knowledge → File Vault menu, in Your Profile › Knowledge, or in Settings › Knowledge in the Studio UI. Wherever a knowledge graph is available, all three choices are offered: Disabled, RAG and RAG + KAG Fusion. Pria keeps a person’s choice: turning Search Only on or off, or turning retrieval off and on again, does not change it. An administrator can set a person’s mode in Admin › Users › Edit › Knowledge under Retrieval Mode (Disabled, RAG Only or RAG + KAG Fusion). Saving that form changes only the settings you edited.

​
KAG Fusion

KAG (the knowledge graph) picks out the people, places, things and ideas each file mentions, and how they relate, and stores that map (a knowledge graph) beside the ordinary index. A question then searches the graph and the ordinary index together and merges the results.
KAG never runs alone. The ordinary search always runs; KAG adds a second search whose results are merged in. There is no “KAG only” mode.

​
Turn it on

The Digital Twin switch also depends on the twin’s account. The note under the switch tells you which case applies: To change the account’s setting, contact the Praxis AI team at humans@praxis-ai.com.

​
Add the files you already have

New uploads join the graph automatically. Files uploaded before KAG was on do not. While KAG is on, a panel under the switch says how many files are not in the graph yet, with an estimate of the segments and tokens to process. Select Add existing files to process them; a progress bar shows the work, and Stop cancels it. Building the graph is an extra processing pass over every file, so it uses more credits than plain indexing. In return it answers questions that thread through several documents better. The AI model that builds the graph is chosen by Praxis AI; it is not a Digital Twin setting.

​
Entity match strictness

In the admin editor’s RAG and Files tab, Entity match strictness (%) (40–75%, default 45%) sets how closely a question must match a topic’s name before the knowledge graph adds that topic’s related facts. Higher is stricter and more precise; lower is broader. Leave it at the default to follow the platform setting.

​
Chunk handling

When a file is uploaded, Pria cuts it into overlapping passages (chunks) before indexing. These settings apply to files uploaded after you save; existing files keep their chunks until they are processed again.

​
Retrieval

These settings apply to every conversation as soon as you save.

​
Showing sources

Display RAG/KAG Search Details (on by default) shows the file passages each answer used — file name, relevance and a preview — above the answer. Turn it off to keep the conversation minimal, for example when file names are themselves sensitive. Confidential passages stay hidden either way.

​
Uploads and the twin’s vault

​
Embedding model

The Embeddings Generation model (admin editor, Personalization and AI Models › System) turns each passage into the numeric fingerprint used for search. The only model offered is text-embedding-3-small, so you can leave this field alone. Different embedding models cannot be compared, so any change would mean processing every file again.

​
Vault scopes

Three vaults feed every search:
  • Personal vault — files a person uploaded for themselves. Only they see them, in every Digital Twin they use, unless Ignore user personal vault is on for that twin.
  • Digital Twin vault — files uploaded inside the twin. Every member sees them.
  • Account-shared files — files another Digital Twin of the same account shares across the account. Every twin in the account can search them.
In Files, the twin’s vault and the account’s files have their own tabs.