quantity and analysis.
In the “Data” menu there are two perspectives for viewing the same book: “Use”. /admin/usage) is ** Record by call. ** for accounting, disruption and cleaning; “analysis”/admin/analytics) is ** Combined signs ** The data source is a volume log: each underlying model call notes a line, not only the user’s visible chat response, title generation, query routing, compressed summary, memory extraction, embedding, images, review, deep research’s planning and verification and other internal calls are also recorded a line (they also generate costs, and should include user costs).
Report of Usage (Data > Usage)
Each API calls a line, in reverse order, with 50 items per page.
List of tables
| 列 | The contents |
|---|---|
| Calling time | Accurate to second, formatted by interface language |
| The User | Mailbox (ID displayed when there is no mailbox) |
| Related Conversations | the conversation title, click directly to the administrator session to view the page; when the conversation occurs in the workspace, the "space" logo is attached below; ** When the conversation has been deleted, “deleted” is displayed. ** No empty ID. |
| The model | The model name. |
| The channel | The actual service of the upstream channel requested; the orange "pants" logo is attached to the bottom of the channel. |
| Purpose of | See the table below; failed request attached the red "Error" logo next to the purpose, click to see details |
| Imports and Exports | Number of input and output tokens. |
| Costs of | Cost of this call calculated by model distribution price (USD) |
| Operated | Remove the button. |
Purpose of Value
| Purpose of | Meaning of |
|---|---|
| Conversation is | The main model call in the user chat, each upstream call in a single multi-round tool cycle. |
| Pictured | Image Generation (Drawing mode shared with image tools in chat) |
| Incorporated | Vectoral calls for RAG document entries and queries |
| Title Generation / Question Routing / Context Compression | Small internal calls of the task model |
| Memory Extraction / Memory Judgment | Asynchronous task calls in memory systems |
| Reduction of manufacturers | Historical Translator Return Time Conversion Call |
| Images of Words | The words before the painting. |
| Research Planning / Research Verification / Research Cross Verification | Task Model Call for Deep Research Pipeline |
| Moderation is | “Model” model review |
verify | Verification of Audit Model in Audit Model |
The single price calculated by the model editing page configuration: text model input / output / cache reading / cache writing four layers (one price per million token), image model pricing, the currency is USD. Only accounting, not deduction here, credits deduction is another layer (see Users, quota and credits)。
Selection and Statistics
| The Filter | Explained |
|---|---|
| The time range | Last 1 / 7 / 30 / 90 Days (Roll window) |
| The User | Mailbox or ID substring, automatic query after entering 400ms |
| The model | Select a single model or all models. |
| State of | All and only mistakes. |
Two statistical cards above the table reflect real-time ** Current Selection ** The total cost and number of records. Any changes in the filter will return to the first page.
Workspace is
Calls in a shared workspace follow "who sends who pays": members continue to talk in a shared conversation, quota and credits consumption recorded ** The sender. ** Under its own name, the measurement line simultaneously records the space it belongs to and displays the "space" logo under the related conversation. When the team makes a cost-sharing, according to the user filter, what is seen is the real consumption of each person, and the space logo is used to distinguish between personal use and collaborative use.
Deleted records
- In-line garbage cans: remove single records.
- "Remove Filter Results": Eliminate all records in the current filtering order in batch, confirm that the bullet window will show the number of bullets; can not be revoked.
Usage logs are not just reports: after the service is restarted, the user’s quota count and time-effectiveness credits are all re-seeded from this table. ** Recently ** Recordings (still within the quota window) may underestimate the number of times used/credits used by relevant users. Cleaning the historical archive data is not affected, and it is recommended to only delete batches according to the earlier timeframe.
Error Request and Failed Load Log
Errors in the account.
When the upstream call fails (connection error, 401 / 403 / 429 / 5xx, etc.) ** The same note. ** Marked as an error state, tokens, fees, credits are all zero.This makes the number of failures statistical, can be filtered according to "only errors", channel quality problems can be seen at a glance.
The calibration guarantees: ** Errors do not affect the user. ** It is excluded from the quota window, the usage count visible on the user’s side and the analytical dashboard, and failure requests neither consume the user’s free number nor infuse the “called number/active user” into fluids that do not coincide with the token/payment.
| circumstances | counted | Deposit of credits |
|---|---|---|
| Normal completed | 是 | 是 |
| User Intermediate Stop (produced content) | 是 | Part of the produced |
| Upstream reporting error (including failure at the bottom) | 否 | 否 |
| Stopped before sending (quota excess / moderation) | 否 | 否 |
Failure load details (only visible to administrators)
Click on the "Error" logo on the wrong line, and the bullet window displays all the scenes needed for the disruption:
| blocked | The contents |
|---|---|
| The upstream error. | The original error returned by the upstream (state code + response, over-length intersection) |
| Request Method and URL | Method of sending to upstream and complete address |
| requested head | The requested photo. |
| requested body | Examples of the request (JSON format) |
Improve the sensitivity: Authorization All sensitive fields such as API Key, Token, Secret, Password, Cookies are replaced by [redacted] The embedded base64 image data is removed as a positioning statement, the request object is deleted to 128 KB by default and the single value is deleted to 8 KB. ** These details are sent only through the administrator interface. ** End users encountering the same failure will only see a generalized error message and will not see the upstream original text.
Disturbance Common Position: State Screening "Only Error" + model filter, open the error details line by line, copy the request body to curl In the upstream reproduction, you can quickly distinguish between "Channel Key fails / upstream changes the protocol / request content exceeds the limit".
Under the logo.
When the model configures the channel at the bottom, the main channel fails to automatically switch to the bottom and try again. The call at the bottom of the channel has an orange “channel bottom” logo (any sub request cut the bottom of the channel once in a tool cycle, the whole wheel is marked). The “channel bottom” logo often appears, indicating that the main channel is unstable, and the channel is repaired.
Analysis of Data (Data Analysis)
Consolidate the same log into a trend view, all statistics. ** Eliminating errors ** The number of calls, active users and tokens/costs are always consistent.
Windows and Indicators
- ** window of time. **: In the upper right corner, select the last 1 / 7 / 30 / 90 days (roll through the window, the interface supports up to 365 days).
- ** Change of indicators **: Number of calls / tokens / fees, one switch drives all charts on the page; tokens = input + output sum.
KPI card
| card is | Meaning of |
|---|---|
| Total number of calls | Total number of successful calls in the window |
| Total number of cards. | Input + Output total of tokens |
| Total costs | cumulative cost by model distribution (USD) |
| Active Users | The number of successfully called decentralized users in the window |
The Trend Map
Align to the column chart to show the overall movement within the window: Select 1 day window. ** hourly ** Bottle and the rest. 天 Separate barrels; the suspension arbitrary columns show the exact value of the time period.
Top Model / Top User
Two parallel cards each list the most consumed in the window. ** The previous 8 models 与 Previous 8 Users ** (Arranged by cost, the same cost by amount of call):
- One horizontal proportion per line (maximum value relative to the current indicator), right-hand value;
- At the end of the row, a mini trend line (sparkline) is attached, and the trend chart is shared, and the consumption of the model / user is constant or sudden;
- The user bar displays the mailbox and the model line displays the model name.
The typical path of cost abnormal scrutiny: the analysis page finds a user's cost surge, records the mailbox, cuts to the "use" page to filter the user, see which model, what purpose, how many tokens it is using, and check the content of the associated conversation if necessary.
Accelerated operating scenes
| The Scene | Operating path |
|---|---|
| User Feedback “Always Failed” | Screw the user + "only error", open the error details to see the original text upstream; if the line is marked with the "pull-down" logo, it indicates that the main channel also fails |
| The channel failed. | Screening “Only Errors”, see which channel is concentrated in the error line; fix the channel Key or give the model the bottom channel |
| Excessive internal costs | Analyze the "cost" indicator, if the share of task calls such as title generation / compression is high, change a cheaper task model to the global setting. |
| Distribute costs to the team. | Analysis page Top users find large households, use page by user filter by line verification, spatial logo distinguish individual and collaborative consumption |
| Free payments are suspected. | Remember that the error line and the intercepted message are not counted; the actual consumption is based on this page record, and the statistics are filtered by user + time range. |
Online user judgment
Online status is displayed in the User > User list: There is a status point in front of each line of the mailbox, green for online and grey for offline; the user information pop-up also shows the precise time of "Recent Activity".
- The User ** In the last 5 minutes. ** Applications that have been issued arbitrarily with authentication are considered online.
- The server updates the user's most recent active time when processing requests to control the frequency of writing. ** Updated every minute. ** Therefore, online judgment has a maximum of about 1 minute of delay.
- The judgment is based entirely on request touch, no independent heartbeat channel: the user opens the page but for a long time without any action (no message, no page) will gradually appear as offline, which is expected behavior.
The "online" list of users is ** This moment ** Nearly real-time status; the "active user" in the analytical dashboard KPI is ** The Statistical Window. ** The number of de-weights generated by the model call. The former answered "who is now in", the latter answered "who is in use during this time."