Recent context
Keeps the latest messages of the current session, serving quotes, follow-ups and continuous conversation first.
Internals
How messages, models, tools and memory fit together
Platform adapters only send and normalise. The assistant runtime owns the decisions, models and tools run inside explicit permissions, and SQLite holds recoverable state.
Architecture
Core modules and who owns what
OneBot v11reverse WebSocket
Telegram / QQ / DingTalkoutbound connections
Feishu / WeComplatform callbacks
ConsoleREST + SSE
Assistant Runtimeevent normalisation · group policy · reply routing · concurrency · replying to the source channel
LLMproviders and role routing
Agenttools, MCP, skills
Memoryhistory, summaries, long-term facts
Pluginssearch, parsing, OCR, voice
SQLiteconfiguration and durable state
Mediaoriginals and history cache
App logsoperations and call chains
| Directory | Responsibility |
|---|---|
cmd/webui/ | Process entry point, environment assembly, SQLite and runtime dependency injection. |
model/assistant/ | Message events, channels, reply decisions, agent, media, memory and tasks. |
model/llm/ | Unified requests, provider adapters, model catalogue, token budget and retries. |
model/storage/ | Schema and persistence for configuration, messages, events, memory and tasks. |
webui/ | Gin API, authentication, configuration, events, updates and SSE. |
frontend-next/ | The Vue 3 admin console. |
Pipeline
How one message decides between replying and staying quiet
- Receive and normalise
The platform adapter turns text, mentions, quotes, images, files and sender details into a single
MessageEvent. - Enrich and cache
Quotes, forwards and media references are resolved, and short-lived platform images are written into the local history media cache.
- Deterministic gates
Checks whether the bot and group are enabled, the reply window, muted users, member level and bot-to-bot suppression.
- Trigger detection
Direct messages are handled directly; groups look for explicit mentions, quotes, trigger words and plugin commands. An explicit mention outranks a quote.
- Semantic routing for unprompted replies
Ordinary group messages that were not directly triggered are judged immediately by the intent model, waiting at most 60 seconds so nothing queues forever.
- Assemble context
Combines the system prompt, relationships, long-term memory, recent history, images and tool evidence, trimmed to the token budget.
- Model and tool loop
Selects the chat or vision role; the agent runs the permitted tools and records every step.
- Send and audit
The reply goes back to the source channel, and the detailed reason, send result, latency and tokens are all persisted.
Memory
Long histories are not stuffed into every request
Compacted summaries
Once history passes a threshold, topics, conclusions and open items are distilled to cut repeated tokens.
Long-term facts
Stable facts, preferences and relationships are extracted asynchronously with their source, confidence and update time.
Retrieval on demand
Searchable by time, keyword, user, group and platform; cross-group and cross-channel access are gated by separate switches.
Retrieval narrows with structured filters first, then ranks by textual relevance and time decay, and finally passes the model only the matching evidence. Summaries are not recomputed when configuration reloads — they expire only when the input history changes or the summary version is bumped.
Multimodal
Originals, OCR and history descriptions each have a job
platform media reference→download and verify→original cache→vision model / OCR→answer and description cache
- Ordinary images are converted into whatever the provider accepts; the media layer does not shrink originals unconditionally.
- Text-heavy screenshots use both a high-detail vision input and OCR evidence.
- History image descriptions are cached by content SHA-256 and algorithm version.
- Image count and vision tokens are budgeted so base64 history cannot grow without bound.
- Video contributes a limited number of context frames, and temporary files are cleaned up afterwards.
Storage
One SQLite database plus a manageable media cache
| Data | Default location | Backup requirement |
|---|---|---|
| Configuration, credentials, sessions, messages, memory, tasks | data/diana.db | Must be backed up, keeping the main database consistent with WAL/SHM. |
| Inbound media | data/media | Can be evicted by size; back it up when historical image recognition matters. |
| Runtime logs | logs/diana.log | Rotate regularly; never publish platform events or debug context. |
| Install and update backups | .installer/backups, .diana-updates | Keep the last known-good version and the evidence needed to restore it. |