Spaces:
Sleeping
Sleeping
Commit History
-
Fix live gaps: auditor copy, direct Ask answers, Ask metadata UI, audit tax detection
Aissugur Kurganbekov +1: Claude Opus 4.8- Landing/Auth/Overview copy -> evidence-backed finance auditor positioning (index.html title+meta, AuthScreen, Overview hero eyebrow/body/footer). - New extractors.py: deterministic regex extraction of common finance fields (total, due date, terms, tax, invoice no, late fee, overdue, uptime). Ask now returns a crisp grounded answer ("The total amount due is $2,413.98.") and falls back to focused-span extraction; values are verbatim, citations unchanged. - Ask metadata panel (question type, evidence, confidence, retrieved/cited, latency) rendered below each answer. - Audit: detect structured fields via the same extractors and fuzzy clauses via keyword scan instead of the retrieval gate -- fixes "No tax / VAT found" on "Tax (8.25%) $183.98" and a parallel false-negative on the contract liability clause. Audit is now fully deterministic. Backwards compatible (Ask, audit, evaluate, auth, scoping unchanged in shape). Verified: 59 backend tests pass, frontend builds, evalua…40c5ba0 -
Add Audit Mode: evidence-backed checklist audit for finance documents
Aissugur Kurganbekov +1: Claude Opus 4.8New POST /audit runs a per-type checklist (invoice / contract / payment_note / general) over a document, reusing the /ask retrieval + abstention gate so it never invents a value. A check is SUPPORTED only when retrieval does not abstain AND the cited excerpt contains the field's discriminator keyword; otherwise the field is flagged unsupported (a missing field drives the risk level). Returns a summary, risk_level, per-field findings (claim + evidence + citation), and metrics (counts + citation coverage). Reuses _owned_document for per-user scoping + 404; /ask is unchanged. Frontend: new Audit tab — a B2B audit dashboard (document + type selectors, risk badge, metric cards, per-finding evidence with highlighted citations). Backend: app/audit.py, audit schemas, POST /audit, and tests/test_audit.py (7 tests: field detection, unsupported detection, citation coverage, false-positive guard, 404, cross-user scoping, default type). Hardened after an adversarial review: field discriminators …f22b512 -
Fix CI: make tests hermetic and harden hybrid_rank against attr-less chunks
Aissugur Kurganbekov +1: Claude Opus 4.8CI installs fastembed (requirements-dev -> requirements), which activated the dense retrieval path during tests and surfaced two issues: - hybrid_rank read chunk.embedding directly, crashing on the SimpleNamespace test stand-ins (and any chunk-like object without the attribute). Use getattr(c, embedding, None) so such a chunk simply stays lexical. - The suite would download the ~130 MB embedding model on every run (slow and flaky). conftest now disables embeddings, so tests are hermetic and deterministic; the TF-IDF+vector fusion is already unit-tested with mocked vectors in test_embeddings.py. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>03863fd -
Add hybrid retrieval (TF-IDF + embeddings), OCR, A/B eval, and answer trust/UX
Aissugur Kurganbekov +1: Claude Opus 4.8Smarter retrieval: - Hybrid dense+lexical retrieval: fuse TF-IDF with fastembed vectors via Reciprocal Rank Fusion (retrieval.hybrid_rank), wired into answering. The abstention gate still runs on TF-IDF so grounding stays conservative. Graceful fallback to pure TF-IDF when fastembed/the model is unavailable. - Store chunk vectors on Chunk.embedding (new column); idempotent ensure_schema() migration adds it to existing SQLite DBs on startup. - OCR fallback for scanned/image PDFs (PyMuPDF + pytesseract), graceful when the Tesseract binary is absent; tesseract-ocr added to the Docker image. - A/B eval harness (app.ab_eval / eval_cli --ab) comparing TF-IDF vs hybrid. Answer trust / UX: - answer_question now returns answer metadata (confidence, question_type, top_score, term coverage, chunk counts, latency); AskResponse.metadata and Citation.match_text added. - Frontend: confidence meter, exact source-span highlighting, and a printable one-page evidence report. Verified: 4…f65a6b3
-
Audit-driven hardening + Google sign-in
Aissugur Kurganbekov +1: Claude Opus 4.8Security (from a multi-agent audit): - CRITICAL: never sign JWTs with a shared constant. AUTH_SECRET_KEY is required in production and a random EPHEMERAL key in dev, so a default deploy can't run on a public, forgeable secret. - Fix cross-user history leak: QA rows now carry user_id and /history is scoped to the asker, so questions over shared demo docs no longer leak across users. - Require auth on /evaluate and /demo/seed (resource-amplification guard). - Guest accounts use CSPRNG creds (crypto.getRandomValues) independent of the email; add a token-type claim; guard citations against a deleted parent doc. Feature: "Continue with Google" (Google Identity Services ID-token flow): - POST /auth/google verifies the ID token (google-auth: aud + iss + email_verified) and get-or-creates the user; password_hash is now nullable for OAuth-only accounts. Gated on GOOGLE_CLIENT_ID (button hidden when unset); /health exposes google_client_id. UX/a11y: honest "session expired" messa…90adb95 -
Add user accounts: registration, login (JWT + bcrypt), per-user documents
Aissugur Kurganbekov +1: Claude Opus 4.8Backend: - User model + nullable owner_id on Document (NULL = shared demo/eval docs). - security.py (bcrypt via passlib, PyJWT HS256) and auth.py (/auth/register, /auth/login, /auth/me + get_current_user). Generic 401 with a dummy-hash compare on miss (no user enumeration), 409 on duplicate, EmailStr validation. - /documents*, /ask, /history now require a Bearer token and are scoped to the user's own docs + shared demo docs; cross-user access returns 404 (never leaking ids). /health, /evaluate, /demo/seed stay public. AUTH_SECRET_KEY is env-driven (prod hard-fail, dev default + warning). - conftest authed-by-default so the 12 existing API tests are unchanged; new test_auth.py (register/login/me/protected/per-user isolation). 38 tests pass, live end-to-end verified. Frontend: - Split-layout AuthScreen (sign in / sign up toggle, validation, show/hide password, "Continue with the demo" guest path) in the light theme. - api.js routes every request through a Bearer wrapper …165a274
-
Harden FlashRank reranking (review-driven) and wire it into the image
Aissugur Kurganbekov +1: Claude Opus 4.8Adversarial multi-agent review follow-up to the second-stage reranker: - answering.py: pool_k = max(TOP_K, RERANK_CANDIDATES) so a mis-set RERANK_CANDIDATES can never shrink the abstention gate's top-K (the abstention decision stays provably on the TF-IDF top-K). - reranking.py: guard the one-time, network-bound ranker construction with a double-checked threading.Lock — concurrent first /ask requests no longer race into two parallel model downloads; bounds-check the result->chunk remap so a changed flashrank result schema can't IndexError. - Dockerfile: flashrank ships via requirements.txt; keep the optional LLM SDKs (openai/anthropic) explicit. Verified: 28 tests pass; golden eval unchanged (7/7, 100% citation coverage, 0 unsupported, evidence-match 1.0); proved enabled-without-flashrank cites the identical set as a no-rerank build, so abstention and eval are preserved. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>1b671d1 -
Add two-stage retrieval: TF-IDF -> optional FlashRank reranking
Aissugur Kurganbekov +1: Claude Opus 4.8Adapts the hybrid-search RAG pattern from awesome-llm-apps (Apache-2.0) as a second retrieval stage. A small cross-encoder (FlashRank, ms-marco-TinyBERT, ~3 MB CPU) reorders the TF-IDF candidates so the most relevant evidence is cited first and fed to the LLM first. - reranking.py: lazy, cached ranker with graceful fallback — if flashrank or the model is unavailable it returns the TF-IDF order, so the app never breaks and still runs fully offline. - answering.py: pull a larger TF-IDF candidate pool when reranking is on, but keep the abstention DECISION on the original top-K, so behaviour and the golden eval are unchanged whether or not the reranker is present. - config: RERANK_ENABLED / RERANK_CANDIDATES / RERANK_MODEL (env-tunable). - Tests for the disabled/unavailable fallbacks and the reorder path (28 pass, golden eval still 100% citation coverage / 0 unsupported). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>ba9f2db -
Add single-container build for Hugging Face Spaces deploy
Aissugur Kurganbekov +1: Claude Opus 4.8One image where FastAPI serves both the built React app and the API, so the app runs on a single port with one public URL (no CORS, no laptop). - app/server.py: parent app that mounts the existing API under /api and serves the Vite build (copied to ./static) at /. Tests, local dev, and the existing two-container docker-compose are untouched. - Dockerfile (repo root): multi-stage — build the frontend, then a Python image that serves it via app.server:root on port 7860, as UID 1000. - README front-matter (sdk: docker, app_port: 7860) so HF Spaces builds it. - .dockerignore + gitignore the local backend/static build artifact. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>81e0253 -
Harden uploads, report real LLM availability, and support free OpenAI-compatible providers
Aissugur Kurganbekov +1: Claude Opus 4.8- Bound the upload read in 1 MB blocks so an oversized body can't be fully materialised in memory before the 10 MB limit returns 413. - Replace the dead ANSWER_MODE field in /health with default_mode + llm_available + llm_provider, so clients can tell whether Thinking mode is actually usable instead of silently falling back to Fast. - Add OPENAI_BASE_URL so Thinking mode can target any OpenAI-compatible provider (free tiers such as LLM7 / Cohere / Mistral) with no code change; document the options in .env.example. - Add tests for oversized-upload rejection and the health contract; make the Thinking fall-back test hermetic and add LLM-answer and LLM-abstain cases (25 passing). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>7269eb9
-
Add Fast / Thinking answer modes (per-request LLM toggle)
Aissugur Kurganbekov +1: Claude Opus 4.8 (1M context)- Fast: extractive engine — instant, offline, free (no API key). - Thinking: LLM reasons on the fly with Claude adaptive thinking (thinking={"type":"adaptive"}, no temperature — correct for current models), grounded in the same retrieved context with citations + abstention. Transparently falls back to Fast when no API key is set, so it never breaks. - AskRequest.mode ("fast"|"thinking", validated -> 422 on bad value); /ask returns the engine that actually answered. - Frontend: segmented Fast/Thinking toggle in the Ask view; mode passed to API. - Thinking model default = claude-sonnet-4-6 (supports adaptive thinking). - 21 tests pass incl. new thinking->fast fallback test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>f2f300a
-
Recruiter-demo polish: curated demo questions, reset-db, PDF test, CORS
Aissugur Kurganbekov +1: Claude Opus 4.8 (1M context)- AskPanel: 3 curated demo questions (strong cite / multi-evidence / abstain), auto-selects the sample invoice for a deterministic 2-click demo - Add `python -m app.reset_db` + `make reset-db` for a clean slate - Add dependency-free PDF upload test (real pypdf parse path) -> 20 tests - CORS: allow any localhost/127.0.0.1 port (and Docker) via regex - README: reset-db command + no-make fallback Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>49eecdc -
FinanceFlow AI: source-grounded RAG Q&A over finance documents
Aissugur Kurganbekov +1: Claude Opus 4.8 (1M context)Full-stack MVP: FastAPI + SQLite backend, React/Vite frontend. - Upload PDF/TXT -> chunk -> TF-IDF retrieval -> focused extractive answer - Two abstention gates + citations for anti-hallucination - Golden-dataset evaluation (UI + CLI) with coverage/abstention/evidence metrics - One-click demo mode, sample-question chips, clean B2B dashboard - Logging, global exception handler, validated/typed API with Swagger docs - 19 pytest tests + GitHub Actions CI - Dockerfiles, docker-compose, Makefile Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>fc5156b