Skip to content

v2.0.0

Latest
Latest

Choose a tag to compare

View all tags
@doeunyy doeunyy released this 24 Oct 23:29
v2.0.0
c7e3f0c
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Veridex v2.0.0 is the version submitted for CSCI 599 Assignment 2: a multi-agent research assistant with RAG, hybrid LLM routing and streaming traces. It runs in its own Google Cloud project; Assignment 1 (v1.0.1, veridex-staging) is unchanged.

Live deployment

Item Value
URL https://veridex-v2-596419356898.us-west1.run.app
Cloud Run service / region veridex-v2, us-west1 (GCP project veridex-v2)
Serving revision veridex-v2-00007-w4f (100% of traffic)
Image us-west1-docker.pkg.dev/veridex-v2/veridex/veridex:v2.0.0
Image digest sha256:eb050ee78f65895a47a067ec3c571646214459294a8a8f78804abf3d4abbb9e9
Source commit c7e3f0c (this tag)
Instances 1–3 (one kept warm during grading, so there is no cold start)

What's included

  • LangGraph Supervisor routing each turn to three specialized agents (Research, Summary, Data), each with its own prompt and MCP tools; a failing agent is recorded and never crashes the graph
  • Hybrid cost/capability router across OpenAI gpt-6.1-sol and Google gemini-3.8-flash, with every decision (rule, reason, signals, Gemini thinking budget) shown in the UI and trace, and automatic fallback to the other provider
  • RAG over 42 CC BY / CC0 arXiv papers (1,855 chunks, Chroma, text-embedding-3-small) with citations linked to source chunks
  • Tavily Remote MCP for live web results; search_evidence and list_documents shared between agents over MCP
  • POST /api/chat/stream Server-Sent Events and a vanilla HTML/JS page with routing chips, an orchestration trace and citation chips
  • Multi-turn conversations stored in Firestore, surviving restarts and new revisions

Results

On 35 fixed questions: RAGAS faithfulness 0.90, answer relevancy 0.79, context precision 0.96, context recall 1.00; $0.024 per question on average. Warm requests show the routing decision after about 2 s and the first answer token after about 11 s. Details in docs/evaluation.md.

Notes

  • The corpus PDFs and Chroma index are not attached; rebuild them with the commands in docs/corpus_v3.md. Attribution is in CORPUS_LICENSE_NOTICE.md.
  • Credentials are never included in the source or this release.
  • Setup, local run and deploy instructions are in README.md; architecture in docs/architecture.md.
Assets 2
Loading