Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.praxis-ai.com/llms.txt

Use this file to discover all available pages before exploring further.

Every Digital Twin has a knowledge layer — the files users have uploaded, indexed for retrieval and surfaced into each conversation as needed. This page is the Admin-side reference for tuning that layer: how much context the model sees, whether the knowledge graph leg runs, what users can and can’t upload, and how retrieval sources are presented in the chat. For the end-user experience — picking Disabled vs. Normal vs. KAG Fusion vs. Search Only on a per-conversation basis — see Knowledge Modes.

​
Retrieval Mode Defaults

The Digital Twin’s default retrieval behaviour is what every new conversation starts with. Users with a knowledge picker can override it on a per-chat basis; users without can’t.
The default. Each user turn triggers a vector search across the in-scope files, and the top matches are spliced into the prompt for the LLM to use. Fast, predictable, and the right choice for most instances.
Runs both a vector search and a knowledge-graph traversal, then fuses the two ranked lists. Surfaces relationships the vector index alone misses (“Who reports to whom?”, “Which course covers what topic?”). Requires KAG to be enabled and incurs extra processing during file upload (see below).
Returns raw search results without an LLM rewrite. Useful when the Digital Twin is wired as a discovery surface rather than a chat — the user gets matched passages with citations and decides what to do next.
Skips retrieval entirely. The LLM answers from its training and the conversation alone — no file context. Choose this when the Digital Twin is purely conversational and uploaded files are out of scope.

​
KAG Fusion

KAG (Knowledge-Augmented Generation) extracts entities and relationships from each uploaded file at ingestion time and stores them as a graph alongside the vector index. At query time, the graph leg runs in parallel with the vector leg; the two ranked lists are merged with Reciprocal Rank Fusion. KAG is experimental and is currently controlled by the Praxis AI team. To request access on your tenant, contact the Praxis AI team at humans@praxis-ai.com to request access.
KAG never runs alone — the vector leg always fires, KAG just adds a second leg whose results are fused. There is no “KAG only” mode.

​
KAG Analysis Model

When KAG is enabled, the KAG Analysis Model field selects which LLM extracts entities and relations from each chunk during ingestion. Lighter models are cheaper and faster; larger models extract richer graphs. The platform default is tuned for cost; institutions with dense documents (org charts, regulations, course catalogues) often see better recall with a larger model. Leave the field empty to use the platform default. See AI Models for the full list of supported models.

​
Chunk Handling

When a file is uploaded, Pria breaks it into overlapping chunks before embedding. Three institution-level settings tune what the LLM ultimately sees at query time.
The number of retrieved chunks spliced into the prompt for each user turn. Default: 2. More chunks = more context the model can ground in, but also more tokens consumed per turn and more chance of distraction by tangentially related passages. Bumping this to 4 or 5 helps long-form research instances; staying at 2-3 is right for chatty assistants.
When on (default), each chunk is augmented during indexing with short keyword summaries that include nearby document context. The embedded text is therefore “this passage about X, situated in a document about Y” rather than the raw passage in isolation. Improves recall on documents where the same phrase appears in many sections. Turn off to embed raw chunks only (faster ingestion, slightly lower recall).
When on (default), chunks are passed through a lightweight LLM rewrite during ingestion to clean OCR artifacts, garbled control characters, and line-break noise from scanned PDFs. Significantly improves embedding quality on image-PDFs and scanned scans. Turn off if your files are already clean text and you want the fastest possible ingestion.

​
Similarity Threshold (kmeanScore)

kmeanScore is the floor for retrieval relevance — chunks scoring below this value are dropped before reaching the LLM. Default: 0.45.
  • Raise the floor (e.g. 0.55+) to be stricter: fewer chunks reach the model, but they’re more on-topic. The LLM is more likely to answer “I don’t know” when the question isn’t covered by your files.
  • Lower the floor (e.g. 0.30) to be lenient: more chunks reach the model, including loosely related ones. Useful when your corpus is small and you want every related passage to be considered.
The default sits in the middle deliberately. Tune it after observing real conversations — if the Digital Twin keeps refusing legitimate questions, lower it; if it keeps wandering into off-topic passages, raise it.

​
Showing Retrieval Sources

The displayRagSearchDetails toggle controls whether users see the retrieval sources panel under each Digital Twin answer.
  • On (default) — every answer that used retrieval shows a small expandable panel listing the source files and the matched passages. Builds user trust (“I can see where this came from”) and supports verification.
  • Off — the answer stands alone; users don’t see which files were consulted. Choose this for instances where document names themselves are sensitive.

​
File Upload Policies

Several institution-level settings govern who can put files into the knowledge layer and how much.
When on, users can chat with the Digital Twin but cannot upload files. The instance becomes read-only at the knowledge layer — only Admins (or users on the email allow-list, below) can add to the corpus. Pairs well with curated institutional knowledge instances where the document set is governed centrally.
A list of email addresses that can upload even when uploads are disabled for users in general. Use this to give specific power users (TAs, curators, content authors) write access to the knowledge layer without opening it to everyone.
The maximum number of files a single user can keep in their personal vault. Default: 300. Prevents one user from filling the vault with thousands of small files. Increase for power users (researchers, content authors); leave at the default for typical learner instances.
When on, the Digital Twin ignores users’ personal-vault files at retrieval time — only files filed under the instance or its parent account are searched. Users can still upload personal files (and use them on other Digital Twins they belong to), but this Twin treats only the curated shared vault as its knowledge base. Improves predictability for instances where the Digital Twin’s answers must be grounded in institutional content only, not whatever the user happened to upload.

​
Embedding Model

The Embeddings Model field selects which model converts each chunk into a vector for retrieval. The platform default (text-embedding-3-small, 1536 dimensions) works well across most workloads. Switch to a larger or domain-tuned embedding only when you have a measured retrieval-quality reason to do so — different embedding models produce non-comparable vectors, so changing the embedding model requires re-ingesting the entire corpus. See AI Models for the supported embedding providers and dimensional notes.

​
Vault Scope Settings

Pria distinguishes three vault scopes that flow into every search:
  • Personal vault — files a user uploaded for themselves; visible only to them, across whichever Digital Twins they use (unless ignorePersonalVault is on for that Twin).
  • Instance vault — files uploaded inside an institutional Digital Twin; visible to every member of that instance.
  • Account-shared vault — files marked account_shared from a sibling instance under the same parent account; visible across every instance under that account.
For most instances the defaults are right: personal + instance + (when applicable) account-shared all contribute concurrently to a single retrieval call. The two settings to revisit are ignorePersonalVault (cover above) and the account-shared flag on individual files (set at upload time).
  • AI Models — embedding model, KAG analysis model, conversation model selection.
  • Configuration — the Digital Twin configuration page where these knobs live.
  • Knowledge Modes — the user-side knowledge picker that pairs with your defaults.
  • Managing Files — how users upload, organize, and share files into the knowledge layer.
  • File Processing Status — what users see while a file is being ingested and embedded.