Histories

The Histories page is the complete log of every conversation users have had with your Digital Twin. It is what you reach for when answering questions like “What did the model actually say to this student?”, “Why are we burning more tokens this week than last?”, or “Show me every conversation that touched topic X.”

​

Overview

Every user message, every Pria response, every tool call, and every retrieval source used to ground a response is captured as a History record. Records are tied to the user, the Digital Twin, the assistant (when one was used), the model that produced the response, and a timestamp. Together they form a conversation log (after 180 days a turn’s tool calls, retrieved passages and reasoning are removed; the conversation itself, models, tokens and credits stay) that powers:

• Audit — answer “who said what, when” for FERPA / GDPR access requests
• Debugging — reproduce a bad response by reading the exact context the model received
• Analytics — top users, most-frequent topics, model splits, error patterns
• Quality work — combined with Feedback, histories close the loop on what to improve

​

Filtering and search

The filter bar at the top narrows the list. All filters compose; selections persist across pagination so you can sweep a large window without losing your place. On top of the toolbar filters, most columns on the Records table have their own filter box — Assistant, Course, the Input / Output columns and more (not Vote, which you can sort). Column filters search only the rows loaded so far (100 at a time; scroll for more), so they narrow a list rather than find every conversation on a topic. Use the Columns button to show or hide columns; ticking Details adds Duration, Credits, Tokens, Discount, Model, Tools, Budget and Claims. Budget says why a tool-heavy answer stopped early (for example “Cut off after 20/20 tool calls”) and Claims flags replies that state an action no tool performed, so you can review them; both have filter boxes of their own. Combining filters is where the page earns its keep: “all conversations on Digital Twin X over the last 7 days”, sorted by Vote, is a quick list of regression candidates.

​

Drilling into a conversation

Click the green pencil button on any row (needs histories.edit) to open the View History viewer: The viewer shows the full message thread end-to-end:

• User turns — exactly what the user typed (or said, in voice mode)
• AI responses — the model’s reply, including any in-line citations
• Tool calls — when the model used a tool (web search, image generation, etc.), the call and its result are shown inline while the Digital Twin’s Display Tools Details setting is on
• Retrieval sources — when grounding was used, the files and the passages that scored highest are listed (while Display RAG/KAG Search Details is on)
• Model metadata — which model produced each response and the token counts

For voice conversations, the transcript is captured in text form so the conversation viewer remains a single, searchable surface across all modes. In a conversation several people share (a Google Chat thread, say), the admin viewer lists every message under the conversation’s owner; open the Secure Link view to see each person’s name next to their messages.

​

Summarize

Pria can summarize the questions people asked into key discussion points, prevalent topics and engagement patterns. Select the Digital Twin, narrow the date range, then click Summarize in the Records toolbar; it reads the 100 most recent matching records and opens the AI Summary panel, where you can ask a Follow-up Question or Print. The summary surfaces:

• The most-discussed topics across the selected window
• Common questions that recurred across users, and outliers
• Profanity or offensive content, with who sent it
• Usage and engagement tables (the top users)

Only the questions are summarized, never the answers. Use this regularly (weekly is a good cadence) to spot drift early — a topic that keeps recurring is often a sign that the Instructions or Files need an addition.

​

Usage and Charts

Two companion tabs give quantitative cuts on the same underlying data:

• Usage — a consumption table per Digital Twin (credits, duration, tokens, queries), broken down further with Show Courses or Show Users, with its own column selection and CSV export; below it, a Peak Usage Hours heatmap and a model-distribution chart
• Charts — a timeline of tokens, queries, credits or users per Digital Twin, combined or compared, by month, week, day or hour; plus Subagent efficiency for the last 31 days

• Filter by date range to compare week-over-week or term-over-term
• Drill into a single user to understand individual engagement
• Export to CSV for sharing in slide decks or external reports

​

Exporting

Two flavours of export, each useful for a different purpose:

• Single conversation — from the conversation viewer, copy the full dialogue (or a link to it) to the clipboard. Best for sharing one specific exchange with a colleague or attaching to an incident ticket.
• Bulk export — CSV downloads the loaded rows with the columns you have visible; Export all prepares the whole filtered list on the server and emails you a download link. Best for spreadsheets and downstream analytics pipelines.

For very large exports across several Digital Twins, contact the Praxis AI team at humans@praxis-ai.com to request a database-level dump.

​

Compaction

Long conversations would otherwise grow without bound and burn tokens on every turn re-sending the entire history. Pria addresses this with compaction:

• Once a conversation grows past a threshold, an LLM summarizes the older turns into a compact representation
• Subsequent turns send the compact summary plus the most recent messages, instead of the entire raw transcript
• The original messages are retained in the Histories database — compaction is a token-saving runtime optimization, not a retention policy
• The compacted summary is cached per conversation so the work isn’t redone every turn; the Digital Twin’s Compact History setting turns compaction off

The conversation viewer always shows the original turns — compaction only changes what is sent to the model at runtime, so auditors and FERPA requests always see the raw conversation, never just the summary. The compaction cache itself is purged after 30 days of inactivity, so a conversation reopened months later will be re-compacted on its next active turn. The underlying messages remain intact.

​

Privacy and retention

Histories contain personal information by their nature — user identity, what they asked, what they were told.

• Encrypted at rest. Database storage is encrypted; access is governed by Pria’s authentication and entitlement layers.
• Admin-only access. histories.list shows the list, histories.edit opens a conversation, and histories.delete deletes records.
• FERPA considerations. When your Digital Twin operates in an educational context, treat histories as covered educational records — restrict admin entitlements accordingly, and use the export to satisfy student access requests.
• GDPR right-to-access. Filter to the user’s email, export, and deliver. The export is the lawful response.
• Right-to-erasure. Deleting a record removes it from the admin pages and from the person’s own history. For hard deletion of conversations to fulfil a confirmed erasure request, contact the Praxis AI team at humans@praxis-ai.com.

Retention windows vary by deployment — see your published Privacy & Educational Consent policy.

​

Using histories for audit

Two common audit flows:

1. FERPA / GDPR access request. Filter by the user’s email, widen the date range (it starts at the last month), export with Export all, and deliver. The export is itself the access response.
2. Incident review. When a user reports that the Digital Twin said something incorrect or harmful, get the timestamp from the user, filter by user + time window, and open the exact conversation. The viewer shows the model, the retrieval sources, and any tool calls — enough to reproduce and root-cause.

Keep an internal log of audit access (date, requester, scope) so the audit trail itself is auditable.

​

Common queries

A few combinations you’ll reach for repeatedly:

• Top users this week — the Usage tab with Show Users, sorted by Queries
• Most-frequent topics — combine the Summarize action with a tight date range
• Model split — the model-distribution chart at the bottom of Usage
• Error patterns — filter the loaded rows’ Output column for “I don’t know”, “I can’t help” or “I’m not sure” to find places the Digital Twin punted; correlate with the Files it should have grounded against

​

Linking to Feedback for the qualitative view

Histories tells you what was said. Feedback tells you what the user thought of it. A History row shows its thumbs rating in the Vote column; a Feedback comment is not linked to a conversation, so match them by person and time. Use them together — a thumbs-down with a comment plus the surrounding conversation gives you the full story.

​

Related

• Feedback — the qualitative companion to this quantitative log
• Sessions — who was signed in when a conversation happened
• Moderation — content moderation policies that may have intercepted parts of a conversation
• Praxis Shield — automatic security-incident detection and triage across conversations
• Privacy & Educational Consent — retention windows and access policies
• Configuration — the Instructions and model settings that shaped every history record
• Assistants — assistant routing visible in the conversation viewer
• Files — the source files that retrieval pulled from
• Entitlements — histories.list / histories.delete permissions