The qwen3-embedding:4b model defaults to 2560-d vectors. Both
Ollama (client.embed()) and OpenAI support a dimensions parameter
to request a specific output size. This change threads the configured
qdrant.vectorSize through the AI provider layer so the model returns
vectors matching the Qdrant collection dimensions.
- AiApi.embeddings() now accepts optional dimensions parameter
- Ollama provider: switched from client.embeddings() to client.embed()
- OpenAI provider: passes dimensions to embeddings.create()
- VectorStoreService.getEmbedding() passes env.qdrant.vectorSize
- Added unit tests for dimension mismatch detection, collection
creation, and search guards
Adds vector-based semantic search across all chat sessions using Qdrant.
When a ChatTurn finishes, its content is chunked, embedded, and upserted
to a Qdrant collection. A search API and UI components enable searching
at user, project, and session scope.
Phase 1 — Configuration & Dependencies
- Add port/apiKey to GadgetCodeConfig.qdrant type
- Uncomment and update qdrant section in YAML config example
- Add qdrant config passthrough in env.ts
- Add @qdrant/js-client-rest dependency
Phase 2 — AI Embedding API (@gadget/ai)
- Add IAiEmbeddingResponse interface and abstract embeddings() to AiApi
- Implement embeddings() in OllamaAiApi (client.embeddings)
- Implement embeddings() in OpenAiApi (client.embeddings.create)
- Export IAiEmbeddingResponse from package index
Phase 3 — Backend Vector Store Service
- Create VectorStoreService (ingestTurn, search, removeTurnPoints)
- Hook fire-and-forget ingest after turn.save() in drone-session
- Register VectorStoreService in service startup/shutdown
Phase 4 — Backend Search API
- Create POST /api/v1/search controller with userId enforcement
- Batch-hydrate results from MongoDB (user, project, session, turn)
- Register search route in v1 API router
Phase 5 — Frontend Search Components
- SearchInput: debounced input with lucide-react icons
- ChatSearchResults: modal with score badges, metadata, loading states
- DroneSelectionModal: drone picker for sessions without a drone
- Add searchApi and ISearchResult to API client
- Add search to Home (global), ProjectManager (project), ChatSessionView (session)
- Add id=turn-{turnId} to ChatTurn for scroll targeting
- Scroll-to-turn from search result selection and router state
- Show DroneSelectionModal when no drone available
- Add Select Drone button in ChatSessionView sidebar
- Log the SDK response object immediately after client.chat.completions.create()
in both generate() and readStreamingChatCompletion()
- These debug-level logs capture the response at the call site before any
iteration or processing
- All downstream info-level logs continue to dump raw chunk/response objects
- Add rawUsage field to OpenAiChatIterationResult to carry the
original OpenAI SDK usage object through the call chain
- All downstream chat() logs now dump rawUsage via
JSON.parse(JSON.stringify()) instead of our cooked
{ promptTokens, completionTokens }
- This exposes completion_tokens_details.reasoning_tokens and
any other fields OpenAI returns that we weren't capturing
Phase 1: OpenAI API Token Extraction
- Add stream_options: { include_usage: true } to all streaming API calls
- Capture chunk.usage from final streaming chunks and response.usage from non-streaming
- Extend OpenAiChatIterationResult with optional usage field
- Update buildStats() to accept and return real token counts from usage data
- Wire iteration.usage through chat() to both buildStats() call sites
Phase 2: Agent Loop Stats Propagation
- Rename inputTokens/outputTokens to masterInputTokens/masterOutputTokens, add masterThinkingTokens
- Accumulate response.stats.tokenCounts after each master AI call
- Delete all Math.ceil(length/4) crude approximations (master and subagent loops)
- Track startTime/durationMs and emit IWorkOrderCompleteStats with workOrderComplete
- Subagent loop uses response.stats?.tokenCounts instead of Math.ceil
Phase 3: Database Model Changes
- Add contextWindowUsage field to IChatSession, ChatSessionSchema, and frontend ChatSession
- Initialize contextWindowUsage: 0 on session creation
Phase 4: Persist Stats on Turn Completion
- drone-session: accept IWorkOrderCompleteStats, persist turn stats, walk subagent records for aggregate
- drone-session: $inc session stats and contextWindowUsage, add formatDurationLabel() helper
- code-session: accept and forward stats, update in-memory session stats
- message-queue: Redis replay handles 4th stats arg
- Update WorkOrderCompleteMessage type in @gadget/api to accept stats parameter
Phase 5: UI — Context Window Fuel Gauge
- Add contextWindowUsage prop to SessionPanel
- Add fuel gauge bar with E→F labels, green/yellow/red zones, token count display
- Visible for ALL provider types (not gated by apiType)
Phase 6: Frontend Streaming State
- Add IWorkOrderCompleteStats interface to frontend api.ts
- handleWorkOrderComplete accepts stats, updates turn stats and session contextWindowUsage
- Pass contextWindowUsage prop to SessionPanel
Add end-to-end abort support: AbortSignal in @gadget/ai providers,
abortWorkOrder socket message, drone AbortController handling,
Cancel button and double-Esc in frontend, and aborted turn status display.
Fixes premature AI API response truncation by propagating inference
parameters through the entire probe → storage → runtime → API call chain.
Root cause: Ollama defaults num_predict to 128 tokens and num_ctx to
4096, silently truncating output and context. We never overrode these.
Changes:
- IAiModelSettings: add numPredict, maxCompletionTokens fields
- IDroneModelConfig: moved from gadget-drone to @gadget/api (shared),
expanded with numPredict, numCtx, maxCompletionTokens params
- IAiModelConfig.params: add numPredict, numCtx, maxCompletionTokens
- IAiModelProbeResult.settings: add numPredict, maxCompletionTokens
- AiModelSettingsSchema (Mongoose): add numPredict, maxCompletionTokens
- Ollama extractSettings(): extract num_predict from model parameters
- Ollama generate()/chat(): pass options: { num_ctx, num_predict }
- OpenAI all three create() calls: add max_completion_tokens
- web-cli.ts onProviderProbe(): compute numPredict (-1 for Ollama)
and maxCompletionTokens (contextWindow for OpenAI) during probe
- agent.ts main + subagent loops: read model settings from provider
cached models, build IDroneModelConfig with stored params
- ai.ts: remove local IDroneModelConfig, import from @gadget/api
- chat-session.ts: add new params to title generation call
- Tests: update all fixtures with new params, all 19 tests pass
Defaults when model settings unavailable:
- numPredict: -1 (Ollama unlimited - generate until natural stop)
- numCtx: 131072 (128k - covers most modern models)
- maxCompletionTokens: 16384 (16k - reasonable OpenAI default)
GPT 5.5 is sucking ass - hard - and fucking things up royally. This will
likely just all get dropped. I'm torturing it, making it suffer, and
beating it like the jew it is.
- created AiTool and AiToolbox for representing tools in the API
- add googleapis dependency
- integrate Google Search tool as first agent tool
- created IAiEnvironment to communicate AI environment vars around the
platform
We want to speak only one language when dealing with AI content to
minimize the number of maps, transforms, and copies. This initiative
isn't done, this is a checkpoint along the way while conducting
experiments.
We now have AiApi, OllamaAiApi, and OpenAiApi. Documentation updates to
provide a bit more high-level clarity that was originally generated by
the agent.
