feat: major update with AI search, deep research, file uploads, chat, and model management#676
Open
SteveLauC wants to merge 265 commits into
Open
feat: major update with AI search, deep research, file uploads, chat, and model management#676SteveLauC wants to merge 265 commits into
SteveLauC wants to merge 265 commits into
Conversation
SteveLauC
force-pushed
the
rebased-chore-ai-search
branch
2 times, most recently
from
27d9a24 to
b4a07c5
Compare
Update the following APIs: 1. POST /settings 2. POST /setup/_initialize Also update the definitions of data sources and assistants. This is currently a mock implementation; the newly added fields are not wired up yet.
…n settings - Add LLMType (string enum) to core/llm_provider.go with constants LLMTypeLanguage, LLMTypeVision, LLMTypeEmbedding; apply to ModelConfig.Type - Add modules/llm/resolve.go: ResolveModel(t, override) reads AppConfig() via modules/common and falls back to DefaultModel when override is unset - Remove mandatory model_provider/model panics from New() in summary, extract_tags, embedding, and file_extraction processors - Relax model_context_length validation to only trigger when explicitly set - Resolve model at every Process() call so settings changes take effect immediately without pipeline restart - Keep embedding_dimension validation (runtime setting, must be in coco.yml)
…del uses
Introduces AssistantModelUse enum (Answering, IntentAnalysis, PickingDoc,
PickingTool) in core/assistant.go and a ResolveAssistantModel helper in
modules/llm/resolve.go.
Each of the four call sites now resolves its model through the helper before
invoking the LLM. The resolution chain is:
assistant config override
-> settings.DefaultModel.<use-specific field>
-> settings.DefaultModel.LanguageModel
Resolved provider/name are written back onto the assistant's ModelConfig so
that downstream readers (PromptConfig, Settings, provider lookup) keep working
without modification.
Note: PickingTool has one extra fallback layer beyond the three-tier chain
above — if nothing resolves, it falls back to the already-resolved
AnsweringModel. This preserves the previous behavior in call_mcp_tools.go
where an unspecified MCP model implicitly reused the answering model, so
existing assistants without a PickingToolModel or LanguageModel default in
settings continue to work instead of failing the MCP tool call.
…id/raw_content/:hint
Adds api.AllowOPTIONSS() to GET /attachment/:file_id and GET /document/:doc_id/raw_content/:hint so the framework auto-registers the matching OPTIONS route. FeatureCORS alone only wires the CORS filter to add response headers; without the OPTIONS route the browser preflight hits no handler and the request is blocked by CORS policy.
Replaces the single POST /setup/_initialize with two public endpoints:
- POST /setup/_initialize — creates the admin user, populates ES
templates (model providers, assistants, MCP servers, roles, etc.),
forces an ES refresh on <prefix>*, then marks setup as done. The
done flag is written last so partial failures leave the wizard
re-runnable.
- POST /setup/_initialize/default_model — sets the default
language/vision/embedding models (creating providers on the fly
when needed) and refreshes the model-provider index. Does not
affect setup completion state and may be called repeatedly.
The UI presents initialization as two steps but the backend only
tracks completion of the first; default-model setup is an independent
operation.
Persists a single KV flag (.setup_done) replacing the prior
.setup_lock. The 'setup' AppSetting now exposes { required: bool } so
the frontend can decide whether to show the wizard.
…ions When the wizard sends multiple selections that reuse the same builtin model provider (matched by ModelProvider.ID) but with different api_token values, the second update would silently overwrite the first. Validate token consistency up front in setupInitializeDefaultModel and return HTTP 400 so the caller can correct the input.
- Add SetupModelDef{id, support_reasoning} to SetupModelConfig.
Users now specify either model_id (existing model on provider) or
model (new model definition), making the two cases explicit.
- core.ModelConfig gains SupportReasoning bool / support_reasoning to
indicate whether a language model supports reasoning mode. Distinct
from ModelSettings.Reasoning which controls runtime use of reasoning.
- ensureSetupModelProvider now receives the LLM type (language/vision/
embedding) so every model written to ES carries the correct Type and
SupportReasoning value. For existing providers the model list is
upserted in place; for custom providers it is included on create.
- Merge validateProviderTokenConsistency + validateModelSelection into
a single validateDefaultModelConfig (5 rules, one loop):
1. Exactly one of model_id / model per selection.
2. model.id required when model is provided.
3. support_reasoning rejected on non-language selections.
4. Custom provider must use model, not model_id.
5. Same builtin provider across selections requires identical token.
Split validation into two receivers on the core types so callers can validate a single model or a whole provider: - ModelConfig.ValidateModelConfig() rejects support_reasoning=true on non-language model types. - ModelProvider.ValidateModels() loops over Models and returns the first error. Replace the package-level helper in modules/llm with these methods. The create and update handlers now call obj.ValidateModels() and return HTTP 400 on a bad payload.
Add a type-level doc comment explaining that ProviderID, Name, Type and SupportReasoning are static identity/capability fields sourced from the provider definition, while Settings, PromptConfig and Keepalive are runtime invocation parameters. Group the fields with section comments for readability.
Add type="language" to every model entry in builtin provider templates and plugin.json files that previously relied on the empty default. Models that already declared embedding or vision keep their existing type. Also drop leftover settings blocks from provider definitions — settings belong to the runtime model invocation, not to the provider catalog. Where a model previously had settings.reasoning=true, promote it to the top-level support_reasoning capability flag.
…ting - Clear provider_id and name in the default assistant store plugin so they no longer hold unresolvable $[[SETUP_LLM_*]] literals that would be written verbatim into ES when the assistant is installed. - Change the RegisterAppSetting key from the nested 'setup' object to the flat 'setup_required' bool, matching the original contract expected by GET /setting/application consumers.
…e dispatch Add a new process_documents processor that routes each incoming document through the enrichment pipeline configured on its datasource. If no datasource-level pipeline is set, it falls back to the globally-configured default pipeline from app settings. Documents with no pipeline configured are passed through to the output queue unchanged. The sub-pipeline is executed synchronously per document so that enrichment processors (embedding, extract_tags, summary, etc.) run before the result is written to the documents_processed queue. Because the sub-pipeline may expand one document into multiple outputs (e.g. a splitter processor), the full result slice is iterated and every enriched message is pushed individually. Wire up the processor in coco.yml inside the process_documents pipeline stage that bridges indexing_documents → documents_processed.
…to document processors Wire DocumentProcessing.LLMGenerationLanguage from the application config into the three LLM-invoking document processors (summary, extract_tags, file_extraction) as a fallback when no processor-level llm_generation_lang is configured. Precedence: processor config > app-level document_processing.llm_generation_language > existing en-US hard default in ValidateAndNormalizeLLMLang(). Also adds the missing "infini.sh/coco/modules/common" import to file_extraction/processor.go required for common.AppConfig().
Validate the incoming document_processing.llm_generation_language value against the BCP 47 standard before merging it into the stored config. Returns HTTP 400 with a descriptive error message when the supplied tag cannot be parsed, preventing invalid language values from being persisted.
Add pipeline.tpl to both zh-CN and en-US setup templates so that the
pipelineconfigv2 ES index is created during POST /setup/_initialize.
Without this, searching pipeline configs before the first pipeline is
created would fail with an index_not_found_exception.
PipelineConfigV2 is defined in the framework and has no explicit
MustRegisterSchemaWithIndexName call, so its index name is derived from
the lowercased Go type name without a -v2 suffix: {prefix}pipelineconfigv2.
Field mapping is handled by the coco-search index template via dynamic mapping.
Add field mappings to the pipeline.tpl setup template so that searching the pipelineconfigv2 index does not fail when the index is empty and dynamic mapping has not yet been populated. Key mapping decisions: - processor: enabled=false to prevent field explosion from arbitrary nested processor configs, while keeping the raw JSON in _source - Standard keyword/date/boolean/integer fields mapped explicitly - labels: dynamic object for user-defined key-value pairs
The .public directory is generated inside the checkout folder (coco/) by vite's outDir: '../.public' relative to web/. The cache path was pointing to the workspace root instead, so cache restore placed the files in the wrong location and subsequent test compilation could not find the infini.sh/coco/.public package.
The unit_test job does not need a real frontend build — it only needs the .public Go package to exist so that main.go compiles. Instead of installing nodejs/pnpm and running a full vite build, generate a minimal .public directory with a dummy file and run make update-vfs to produce a valid static.go package stub.
…bs/coco-server into rebased-chore-ai-search
pnpm 10.x introduced a breaking change requiring packageManager dependencies in the lockfile to carry a registry path and integrity hash. Since the lockfile was generated with pnpm 9.12.3, CI must use the same version. Replace all 'latest' and unversioned pnpm installs with the pinned 9.12.3 to match the packageManager field in package.json.
…bs/coco-server into rebased-chore-ai-search
…bs/coco-server into rebased-chore-ai-search
Add missing documentation for the file_type_detection processor. Remove internal Go type references (core.Document, []queue.Message, struct field paths) from all processor docs so the documentation is oriented toward users configuring pipelines rather than developers reading source code.
Reverts 1c1da75. The lang parameter on the cancel API and the CancelLang field on MessageTask coupled the backend to UI language concerns that belong in the frontend. The cancel flow now simply invokes CancelFunc without persisting a localized message.
…bs/coco-server into rebased-chore-ai-search
…bs/coco-server into rebased-chore-ai-search
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is a large commit containing many significant changes:
Standards checklist