Skip to content

Latest commit

 

History

History
394 lines (286 loc) · 11.9 KB

db-schema.md

File metadata and controls

394 lines (286 loc) · 11.9 KB

Database Schema Design

Note this is a guide, the real design may diverge and this is not currently fully implemented anyway.

  • everything has a db id
  • file references may be done in either direction; either the filename contains the database id or there is a column to capture filename. Large media files belong on the filesystem and ideally their content can be identified or categorised conveniently.
  • consider everything making sense when:
    • sometimes different subsystems are disabled at runtime
    • physical kiosk deployment
    • online video call deployment
    • online audio-only call deployment
    • online remote deployment (e.g. staging/test)

File References

How should media files integrate with database records? Text messages to and from the LLM can be stored directly in the db but the filesystem feels more suitable for images, audio and video files.

Linking rows in the database to files can be done by using the db id in the filename or by storing the filename in the table.

A query to show the primary view of a conversation would start by fetching the session with the sequence of messages for that session. For each message, fetch all output chains; all rows with the message as the input and recursively fetch each link in the output chain using the previous output as its input.

File names with db id

Either the row must be inserted before the file is written so the id is known or the file is written with a temporary name and renamed according to the db id once the row is inserted. If the table is to store file contents metadata like duration for audio and video, this cannot be known until the file is written - at least for cases where the file is written by an external process.

No extra index is required on the file tables to efficiently look up the row for a given file.

Generated filenames stored in db

The db row can be written with a filename column and the filename can be generated independently of the database.

To go from a file to a row, the query might need an index on the filename column to avoid inefficient table scan.

LipSync Combined with TTS

Some services offer to convert text to lipsync rather than being driven by speech audio and therefore do not have an intermediate tts or audio record. The lipsync database record for these products would need to link to a message rather than a tts entry. One way to implement this is to have two foreign key columns on lipsync, one for a tts input and one for a message input such that all lipsync rows should have exactly one of these columns populated and the other should be null.

Because a combined TTS and Lipsync service could have lower total latency than executing the two in series, it seems worth redesigning the current db schema to support this.

system_version

Need to reference code version somehow. Decide if it makes sense to make an upgrade during a scenario or maybe it's fine to make a new scenario. Maybe each scenario has a system_version?

  • version tag - semver
  • git hash - base hash
  • modified - boolean (devmode)

run

Created each time system is booted.

  • started - datetime
  • component - server/frontend/? they can have different lifecycles

deployment_type

e.g. devmode, A/V production event, video/audio call setup

  • descriptor
  • configuration

scenario

Represents a physical scenario

  • created - datetime
  • scenario
  • configuration - various subsystems can be disabled
  • deployment_type - fk

user

Authenticated users can engage in training mode. Can have kiosk users who equate to a deployment situation where multiple people interact under this shared identity.

  • username
  • external identifier
  • other relevant coordinates (if audio or video call, some identifier)
  • permissions?
  • email

calibration settings

Because a camera can move, possibly even mid-session, some features might need to have settings that change accordingly. Autonomous recalibration is possible in-principle, but recording and capturing this change is necessary for quality assurance and improving such automation in future. It might be necessary to interpret pose estimation quality post-hoc and to preserve, correlate or contextualise the information contained in judgements.

session

Identifies a sequence of interactions by a user

  • user - fk - system is used with relevant permissions
  • startTime
  • finishTime
  • timezone - punter timezone (system timezone could be different)
  • clientsystem - fk
  • initiator (workflow trigger mechanism, implicit, explicit etc.)
  • deployment - a session always has a containing deployment

mode_state

During a session, the system will change mode states, these entries record specific state transitions.

  • session - fk
  • start - datetime
  • mode - fk/enum - which mode is this?

client_system

Capture details about the client runtime and hardware if known, alternately capture relevant details if is an online voice or video call

  • browser details
  • camera
  • microphone
  • speakers
  • resolution
  • external account references - if online through some call channel
  • ... etc.

portrait

Character portrait to use for showing and to use for driving lipsync video. May need to do format conversion, resizing etc. Probably best to have one master image and derivatives refer back to original. Consider same idea for other media assets and their transformation pipelines.

  • file ref - need to know which is master image, which are post-processed?
  • image hash
  • width
  • height
  • filesize
  • format
  • other image metadata (bit depth yada)
  • provenance - where did this come from?

video

Could be captured input for pose estimation, generated output from lipsync, used as input for gesture reference etc.

  • created - timestamp
  • video file reference
  • length_ms
  • mime_type
  • post-processing
  • type - input/output
  • provenance - ?

audio

Smallest unit of stored audio

  • created - timestamp
  • length
  • format
  • audio_filter - fk, optional only if applied
  • type - raw input, tts output
  • provenance - where did this come from?

stt

Speech to text

  • created - timestamp
  • system
  • parameters
  • audio_recording - input
  • text - output

prompt_template

Forms a graph but how to resolve/store value references for usage instances?

Templates for producing system prompt text with named holes for parameters that are dynamically filled (e.g. current time, another prompt template)

What format to use?

  • text

prompt_parameter

  • unique_name
  • function_reference - code refe.g. current datetime refers to a function that always returns a different value or a prompt template?

chat

Text to text

  • created - timestamp
  • system - could be composite, expert system, LLM etc.
  • parameters - expert system should betray its options variant
  • system_prompt - fk
  • user_prompt - optional (could be initiator)

modestate

Workflow graph values may not belong in db but need to be referenced: idle, attract, invite, chat, close etc..

Need a "warmup" mode that must complete before user-interactivity. Might cover pregeneration of canned responses or file format conversion. System needs to have well-defined front-end behaviour during each mode. The warmup mode might be a special case.

Maybe a reset mode where the server is inacessible from the front-end and it indicates differently that it has no idea what is going on because it can't reach its server to even tell it if the network is down or that it is busy.

Might need fault modes like if the network goes down unexpectedly. Maybe need to precalculate front-end behaviour to explain this. Such as making the persona go to sleep, meditate or have left the scene leaving an empty chair.

It would be cool to be able to generate a meditating closed-eyed version of each source portrait using pose-estimation and face detection on the source portrait. Also to use image to text to figure out if the image generation succeeded.

These might be deployment specific and be implemented differently depending on available subsystems. If no video, then triggering a mode state transition needs to have an alternative defined mechanism than pose estimation.

tts

  • created - timestamp
  • system
  • parameters - ?
  • audio_recording - fk
  • text - input

lipsync

Generated video from speech audio, portrait and maybe gesture reference

  • created - timestamp
  • system
  • parameters
  • audio_recording - fk - input
  • portrait - fk - input
  • gesture_reference - fk - optional input
  • video

vision

Image-to-text interpretation of image contents for self-image, description and detection of punter or interpretation of arbitrary scene from some other cam.

Will be used as template parameter input for system prompt or embeddings.

  • created - timestamp
  • system
  • parameters
  • image - file ref - not necessarily a self or punter portrait
  • description - basic description of image contents

audio_filter

Some small number of audio processing definitions to help categorise different pre-processing methods and distinguish rows in audio_recording

  • name
  • implementation - code ref?

pose

  • video file reference
  • provenance (synthesised, from a session/deployment?)
  • pose system parameters
  • captured pose coordinates?
  • interpretation by system

pose_sequence

Some kind of chain of poses captured in sequence

  • index
  • pose - fk

embeddings

Not sure how these work yet, but it would bulk load information to be used for reference by the persona. It could be an event schedule for the deployment or some set of predefined lore loaded this way as an alternative to training, fine-tuning or system prompt.

cached_uploads

Online services often require that file uploads are first submitted, from which they provide a globally accessible URL. Then that URL may be referenced as the file input for an API call. Such URLs should be cached in the db so they can be reused rather than a fresh upload of the file be performed every time the same local file is to be used as an input for the call.

FAL sadtalker uploads are a good example of this since the input image is reused for many calls for lipsync.

judgement

Human judgement of a production artifact

  • score - -1 to 1
  • user - fk
  • created - timestamp
  • active - can deactivate
  • comment - optional

judgement joining tables

Also need joining tables for each type that can receive judgement, probably including other judgements, each looking similar.

  • judgement - fk
  • target id - fk specific to table type

performance logs

similar to judgements, may need multiple tables or maybe stored in a different db, in filesystem or some hybrid

  • created - timestamp
  • measure - string key identifying measured action
  • value - always a duration in ms?
  • hardware - physical machine details if relevant to enable comparison
  • system - service identifier
  • parameters - might include model options etc.
  • resource details - sometimes available ram is relevant
  • deployment - fk
  • session - fk
  • system version - fk

faults

All runtime failures should be captured for later analysis and correlation with judgements. Generally these are exception flows due to outages or bugs but can also be detectable unwanted outputs (video format, length, size invariants can be broken). LLM refusal may or may not be considered a fault or a low quality legitimate output?

  • created - datetime
  • type - some short category (maybe also detect soft refusals by models)
  • description - detail
  • system_version - fk
  • deployment - fk
  • session - fk
  • current mode
  • ... other details similar to performance logs

precalculation metadata

Various artifacts should be generated either on-demand or preemptively. We may want to automatically generate combinatoric versions of canned tts and lipsync video automatically when portrait images update. These could be done in a pre-startup baking operation which is more like a modestate. A record of what has been or is to be precalculated might need db tables.

test suite specific data

Hopefully this can be done independently of the database as much as possible, but some warranted automated integration tests may generate production-like