Spaces:
Sleeping
Sleeping
Commit History
-
Fix gpt-5.4-mini tool calls and stuck chat thinking indicators.
NeonClary +1: CursorOmit reasoning_effort when calling OpenAI with function tools, clear thinking state on stream complete, and surface expired JWT so chat no longer silently stalls. Co-authored-by: Cursor <cursoragent@cursor.com>99ae678
-
Wire FEAT auth, bootstrap, vLLM-with-Neon-pile, profile-sync, conversation summary, and avatar/UserUpdate model into upstream/dev base
NeonClary +1: CursorCo-authored-by: Cursor <cursoragent@cursor.com>923dda1 -
Add cybersecurity overlay: configs, cyber personas, SQLite shim, onboarding/profile, Canvas widgets, deployment infra
NeonClary +1: CursorCo-authored-by: Cursor <cursoragent@cursor.com>c6537a8
-
Feat/advisor tools (#38)
NeonCharlie-24Unverified signature* added /scraper directory from Clary branch for inspection. * Added Gemini function-calling MVP with course search tool. * Added Rate My Professor tool and multi-tool dispatcher. * Added tool registry with auto-discovery and tool filtering. * fixed test warning in RMP tool test. * added structured tool-call return result and get_tool_response method to orchestrator. * wired tool calling into orchestrator chat routes. * cleaned up standalone test routes and did minor refactoring. * fixed context manager dropping non-allowlist persona and tool responses. * pinned backend requirements to <pkg>~=M.m from latest container build. * removed license notices and cu_info_scraper. * moved the tool-calling loop into LLMClient with provider specific overrides implemented in each respective client class. * refactored tools config to support per tool settings. * added rmp school_id lookup script for CLI usage. * added dark mode awareness to advisor color fallback which is used by the o…68ee171
-
Fix/json output (#40)
NeonCharlie-24Unverified signature* added json response mime type to generate_contextual_clarification. * added structured JSON output support for vLLM client. * added structured JSON output to get_top_personas and handled dict-wrapped responses.b2e8e99
-
Feat/vllm endpoints (#36)
NeonCharlie-24Unverified signature* added initial tests for vllm client class. * added the ImprovedVllmClient class. * added vLLM option to config, bootstrap, and api routes. * added vLLM option to front-end provider dropdown. * added api key to docker-compose file and removed it from the yaml config files. * added error handling and design decision comments. * removed example vllm client file. * added model re-discovery on request failure. * convert vLLM to a single client and format persona roles to be vLLM acceptable names (system/user/assistant). * moved _clean_responses into LLMClient base class to eliminate duplication. * changed vLLM API label color to red and also set collapsed button color from grey to the respective providers color.be70bd4
-
Fix/structured json output (#30)
NeonCharlie-24Unverified signature* Added structured JSON output for canvas insight extraction. * changed model to gemini-2.5-flash in phd_config.yaml, undergrad_config.yaml, and config.py * fixed canvas insights JSON parsing by disabling thinking to preserve token budget. * Updated LLMClient.generate docstring to document all parameters. * Removed LLM response JSON extraction patching. * Increased max tokens in _extract_insights_from_content from 500 to 1000.484a699
-
Improve backend logic around user input parsing (#12)
Daniel McKnight +1: NeonClaryUnverified signature* fix: improve canvas confidence scores and insight extraction Switch insight extraction from fragile JSON parsing to plain-text numbered list format. Add content-driven confidence variability: LLM-extracted insights score 0.75-0.90 based on length, rule-based insights score 0.55-0.70 based on pattern match count. Made-with: Cursor * fix: improve orchestrator clarification and LLM response handling Rewrite _needs_clarification to check keywords before length, preventing false triggers. Add generate_contextual_clarification using LLM for tailored follow-up questions. Wire clarification check into /api/chat. Filter Gemini 2.5 thinking-mode parts from responses. Clean up _clean_response in both Gemini and Ollama clients to preserve Markdown. Made-with: Cursor * Address issues noted in review * Revert JSON prompting and clean up imports * Refactor input sanitization to log malformed input requests --------- Co-authored-by: NeonClary <askclary@gmail.com>98d14cc
-
Add configuration support (#3)
Daniel McKnight +1: NeonClaryUnverified signature* feat: make app fully configurable via config.yaml and /api/config endpoint Backend: - Add Pydantic-validated config.yaml with all app settings, personas, LLM, auth, MongoDB, RAG, and orchestrator configuration - Expose public (non-secret) config via GET /api/config endpoint - Refactor auth, database, bootstrap, orchestrator, RAG manager, embedding client, Gemini client, canvas analysis, and chat summary to read from config instead of hardcoded values - Persona definitions (names, prompts, temperatures) driven by config.yaml with shared base_prompt + per-persona persona_prompt - Environment variable fallbacks preserved for secrets (API keys, JWT, MongoDB) Frontend: - Add AppConfigContext that fetches /api/config and provides advisors, colors, icons, and all UI text to every component - Update HomePage, ChatPage, CanvasPage, Login, Signup, Sidebar, AdvisorCard, MessageBubble, ThinkingIndicator, SuggestionsPanel, AdvisorStatusDropdown, and chat inputs to consume config c…2ff64f6
-
Response formatting fix
Sohan Kshirsagar9846f7b
-
Backend Documentation Addition
Sohan Kshirsagarbbb4f78
-
updated gemini context management
Girish Jeswanib19def4
-
remove redundant code
Girish Jeswani5ff4a89
-
update advisor name
Girish Jeswani797ce6f
-
context fixes
Girish Jeswania9328c2
-
Merge branch 'main' into sohank-feature
Sohan Kshirsagar599da37 -
reconfiguring llm classes to match the persona changes for response generation
Sohan Kshirsagardab9098
-
Embedding changes
chandanasunkara0d2283a -
Merge branch 'main' into chroma_db
chandanasunkara108073e
-
file upload feature
Sohan Kshirsagar195b80e
-
BACKEND : Setup Database to store and retrieve information
chandanasunkarac6ce494
-
modify response length
Girish Jeswani05e445e -
update llm client for better responses
Girish Jeswani00b17e7
-
gemini changes
Girish Jeswanidc1fb47
-
add generic ollama client
Girish Jeswanie1b8548
-
Changes to include single model represented as multiple personas
Sohan Kshirsagar3bc3e37