Spaces:
Running
Running
Commit History
-
Fix eval judges and publish first scorecard results Running
- Relevance and groundedness judges now see the full retrieved context; they were given truncated excerpts and penalized grounded answers - Correct two reference answers against the docs: full backups are kept 90 days (not 7), and stop-writes thresholds are documented - Add results/eval_results.json: correctness 84%, groundedness 92%, retrieval relevance 87%, out-of-scope handling 100% Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>bb393d8 -
Make escalation a calmer, more selective suggestion
- Triage recommends escalation only for P1/P2 issues needing hands-on engineering, or needs the docs can't meet (billing, account access, data recovery, likely bugs) - not routine how-to questions - Show it as a blue "Suggested escalation" note instead of an orange warning that read like an app error Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>57893d9 -
Compute source highlights on demand; raise default daily cap to $1
- Highlights are now fetched from POST /highlight when a source is opened (cached client-side) instead of embedding every sentence of all 15 retrieved sections per analysis. Retrieval drops from ~2s to ~0.75s per question locally, more on the 2-vCPU Space. - Default DAILY_SPEND_CAP_USD is $1.00, and a blank value falls back to the default instead of crashing startup. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>a0255c7 -
Fix OOM on startup ingest; halve analysis latency; add About modal
- Embed SPLADE 4 chunks at a time during ingest: the default batch ran the HF container past its 16 GB limit (now peaks ~3 GB); allow a 1h startup for the first boot after a docs change - Extract questions and triage in two parallel calls so research starts ~5s in; shorter triage fields and ~150-word answers (29.7s -> 14.6s end to end in local Docker, 5.1c -> 4.3c per analysis) - Reply drafts are plain text with no markdown - Load backend/.env.local for local runs - About modal: how the pipeline works, hosting and budget limits, privacy Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>bdf34b2 -
Rebrand as Halyard Cloud ticket analyzer; add triage, replies and personas
Branding - Replace Aerospike with the fictional Halyard Cloud product across the UI, prompts, eval set and scraped docs; drop source URLs and the crawl index - Rename the HF Space to tomeese/support-ticket-analyzer in both workflows Fixes - Spend cap: always record spend, block the next request once over the cap - start.sh: ingest fallback never ran under set -e; also re-ingest when the docs fingerprint stored in Qdrant doesn't match the docs folder - Run endpoints in the threadpool and answer questions concurrently - Structured outputs for triage, answers and the eval judge - Input size limits; remove the unused /chat endpoint and chat UI - Pin dependencies; embed with fastembed (ONNX) instead of PyTorch Features - Customer persona (account, experience, mood) shapes triage and replies - Triage card: severity, category, sentiment, escalation - Streamed progress, "needs review" flag from documentation coverage - Clickable citations with a source viewer that highlights the passage …d0954af
-
add confidence scores to finding cards
Qdrant RRF scores are averaged across retrieved chunks and normalized against the theoretical RRF maximum (2/61). Displayed as a color-coded badge on each card: High ≥70% (green), Medium 40–69% (amber), Low <40% (red). Supports future promote-to-KB workflows based on retrieval confidence. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>3bcd15a
-
replace chatbot UI with ticket analyzer
- Backend: add /analyze endpoint — extracts 3 questions from a ticket via Claude, then runs each through the RAG pipeline and returns structured findings - Frontend: new TicketAnalyzer component with ticket textarea, 4 sample tickets (VPC peering, backup restore, auth failure, Grafana setup), animated loading states, and per-finding answer cards with sources Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>7c8bea6 -
fix health endpoint and add daily keep-alive workflow
- Fix /health to query Qdrant instead of stale ChromaDB reference - Add keep-alive.yml to ping the Space daily with a 3-min timeout, preventing Qdrant cluster inactivity shutdown Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>73343e4 -
update Claude model to claude-sonnet-4-6
claude-sonnet-4-20250514 has been retired; replace with current model ID. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b24b3a4
-
cleaned up makefile and added start.sh to allow the creds to be pushed down
88629d8 -
removed stale ingest file for chroma.
f2d0099 -
UX updates.
swapped out chroma_db for qdrant. Added w&b1d67421
-
add HF Space sync action
a052b68