Skip to content

session_context.md

Latest commit

 

History

History
54 lines (38 loc) · 4.66 KB

session_context.md

File metadata and controls

54 lines (38 loc) · 4.66 KB
Raw

Session Context — 2026-07-11

Branch: 2.3.0-dev · Repo: F:\Indicragv2\IndicRAG · Plan: F:\Indicragv2\implementation_plan.md

Handoff doc for resuming after compaction. Everything shipped this session is committed; working tree carries only the local .env change (gitignored).

What happened this session

1. Phase 8 (OpenRouter) — planned only, deferred to LAST

Added Phase 8 to implementation_plan.md: OpenRouter as a first-class secondary LLM provider + user-selectable models. Design settled:

  • Provider abstraction behind generate_with_failover / llm_generate_stream; OpenRouter backend returns a Gemini-shaped response shim so the ~6 agent call sites stay unchanged.
  • Curated .env allowlist LLM_SELECTABLE_MODELS (e.g. gemini-3.5-flash,anthropic/claude-haiku,openai/gpt-5.4-nano) drives the UI dropdown; live OpenRouter catalog only enriches capability metadata.
  • LLM_PROVIDER / LLM_FALLBACK_PROVIDER make either vendor primary or fallback.
  • Scoped 6 d → 8 d after folding 6 gaps (re-export constraint at rag.py:577, per-backend error classifiers, (provider,model) circuit key, net-new streaming cross-provider failover, google-genai stays a hard dep, SAFETY_SETTINGS dropped on OpenRouter, pre-commit capability gating). No code written — deferred behind Phases 5–7.

2. Phase 5 (contradiction detection) — VALIDATED end-to-end

Ran the real mDeBERTa NLI model (offline) against the plan's stated "over-firing on paraphrase" risk. Result: opposing claims 0.724, paraphrase 0.019, unrelated 0.126 — threshold 0.6 well-placed, wide separation, no false positives. Ships as-is, no code change. Memory updated. (Real-model check kept in scratchpad; repo tests deliberately mock NLI for offline CI.)

3. Phase 6 (watch-a-topic) — building incrementally

Decisions: manual-trigger first (schedule loop deferred) + full-PDF arXiv ingest (no-PDF → abstract-only). Single-worker, no Redis/Celery.

  • ✅ Increment 1 — f70ae57 — watches table + CRUD in persistence.py (save_watch/get_watch/list_watches/due_watches/delete_watch). Mirrors job/prefs pattern. Tests: tests/test_watch_persistence.py (6).
  • ✅ Increment 2 — b5eb6dd — routes/watch.py CRUD (POST/GET/GET{id}/DELETE /watch), gated by config.WATCH_ENABLE (404 when off). Router wired in api_server.py. Knobs WATCH_ENABLE + WATCH_DEFAULT_CADENCE in config.py + .env.example. Tests: tests/test_watch_routes.py (8).

18 watch tests green; no regressions.

NEXT: Increment 3 — run_watch core + POST /watch/{id}/run

Design (contracts already confirmed):

  1. execute_arxiv_search(topic, max_results) (+ execute_open_access_search fallback) → hits {text,title,source,pdf_url,arxiv_id}.
  2. Filter hits whose arxiv_id ∉ watch.seen_ids.
  3. Per new hit: download pdf_url → temp file → ingest.ingest_pdf(path, paper_id=arxiv_id, metadata={title,source}) → returns (n_chunks, title); n_chunks > 0 = genuinely new (built-in paper_id/content-hash dedup). No pdf_url → abstract-only chunk.
  4. Summarize the newly-ingested papers into a cited digest (LLM).
  5. Update seen_ids += new arxiv_ids, latest_digest, last_run=now, advance next_run = now + cadence; save_watch.
  6. Endpoint: POST /watch/{id}/run — sync via run_in_threadpool (ingest blocks the event loop otherwise).
  7. Tests: mock execute_arxiv_search, ingest_pdf, and the LLM summarizer; assert dedup, seen_ids growth, digest storage, next_run advance.

Then Increment 4 (asyncio lifespan loop over due_watches) and Increment 5 (GET /watch/{id}/digest + UI).

Gotchas / how to test

  • Restart the server after .env changes — config.py reads env only at import.
  • .env has WATCH_ENABLE=true locally (gitignored, NOT committed). /watch returns {"detail":"Topic watches are not enabled"} (404) when off — that's the gate working.
  • Run tests: python -m pytest tests/test_watch_routes.py tests/test_watch_persistence.py -q (needs PYTHONPATH = repo root for standalone scripts; pytest handles it via conftest which also redirects the SQLite DB to a temp file).
  • GateGuard fact-forcing hook fires on first-touch of each file + first Bash. Disable with export ECC_GATEGUARD=off or ECC_DISABLED_HOOKS += pre:edit-write:gateguard-fact-force.
  • No Co-Authored-By Claude lines in commits (user preference).

Watch dict schema

{id, user_id, topic, language, cadence(daily|weekly|monthly), seen_ids:[], latest_digest, next_run(ISO-8601 UTC|null), last_run, created_at} — stored as json in watches.data; user_id/next_run/last_run/created_at denormalized as columns.