Spaces:
Sleeping
Sleeping
Commit History
-
Fix: Groq decommissioned llama-3.3-70b-versatile
Every /chat request returned 500. The Groq API key is still valid, but llama-3.3-70b-versatile is no longer in the served model list, so the LLM call failed immediately. llama-3.1-8b-instant, used as the eval judge, is gone too. - LLM_MODEL -> openai/gpt-oss-120b, JUDGE_MODEL -> openai/gpt-oss-20b - Both read an env override (FINCHAT_LLM_MODEL / FINCHAT_JUDGE_MODEL) so the next retirement is a Space-secret change rather than a redeploy - Pin citation format in the prompt and normalise output deterministically: gpt-oss emits CJK-bracketed citations and typographic punctuation regardless of what the prompt asks for Verified locally: narrative, numeric ratio, YoY trend and the no-answer refusal path all return correctly formatted answers. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>62a7aca
-
Add deterministic computed-ratios layer (Approach 2); FinanceBench 20% -> 45%
087a312 -
Fix stale '18-company' references (corpus is now 25; make eval text count-agnostic)
ca49d1a -
Expand corpus to 25 companies (add Apple, Alphabet, Amazon, NVIDIA, Tesla, JPMorgan, Walmart)
6f598ab
-
Update README + evals for 18-company EDGAR corpus; add capability eval (100%) alongside FinanceBench (20%)
47541b7
-
Expand corpus to 18 companies (recent SEC 10-Ks); ship prebuilt index via git-lfs
7ab3f4b
-
FinChat: RAG chatbot over SEC 10-K filings
6ba3ef3