NotebookLM Research (Gemini long-context)

NotebookLM-equivalent analysis of large documents. NotebookLM has no official public API, so this is built on the Gemini API using gemini-3.1-pro-preview (~1M-token context). Grounds answers in your documents; supports Arabic + English.

When to use

• Summarize / question / extract from large PDFs, txt, md.
• Multi-document synthesis grounded in sources.

Setup

• Env: GEMINI_API_KEY (or GOOGLE_API_KEY). Both present in secrets/env.sh.
• Python stdlib + requests only.
• Optional model override: NBLM_MODEL (default gemini-3.1-pro-preview).

Usage

bash
. /data/.openclaw/secrets/env.sh
Ask (Arabic)
python3 scripts/analyze_docs.py --files report.pdf --ask "ما الفكرة الرئيسية؟" --lang ar
Summarize / Extract
python3 scripts/analyze_docs.py --files a.pdf b.txt --summarize
python3 scripts/analyze_docs.py --files notes.md --extract "dates, names, amounts"
NEW — shareable report (NotebookLM-style)
python3 scripts/analyze_docs.py --files spec.pdf --report --lang ar
NEW — audio overview (spoken summary -> MP3 via OUR TTS arsenal, failover-safe)
python3 scripts/analyze_docs.py --files spec.pdf --audio /tmp/overview.mp3 --lang ar

2026 update + dual-account

• NotebookLM (the product) moved to Gemini 3.5 default + source-repository-from-chat + reports + slide decks (2026-06-08). Still NO public API — this skill stays the Gemini-based equivalent.
• Dual Google accounts (Kuwaiti GEMINI_API_KEY + US GEMINI_API_KEY_US) with automatic failover (never-stop).
• Outputs now include shareable report + audio overview (reuses media_failover.py TTS: Gemini Charon -> xai -> ElevenLabs WAEL -> MiniMax).

Integration with primary RAG (auto-routing)

• Primary RAG (ask.drwaelalkishawi.com) = our private full-corpus recall. ALWAYS primary for internal questions / our own knowledge.
• This skill = COMPLEMENT for quick shareable outputs from a defined source set: polished report, audio overview/podcast, or deck (via slideshow/HyperFrames).
• Auto-rule: internal recall / "what do we know about X" -> RAG + memory_search. "Make a shareable report/podcast/summary FROM these docs" -> this skill. Sensitive docs (family/real estate/KCPC) -> keep local, do NOT send to Gemini.

How it works

• Small text/md (<200KB) → sent inline for speed.
• PDFs / large / many files → uploaded via Gemini Files API (resumable upload, wait for ACTIVE, reference by URI), then deleted after the call.
• --inline forces inline (text only); --lang auto|ar|en.

Verified (2026-05-31)

• gemini-3.1-pro-preview live via v1beta/models/...:generateContent ✅
• Inline Arabic Q&A ✅
• Files API upload (resumable) + PDF Q&A + cleanup ✅

Limits

• True NotebookLM features (audio overview, notebook UI, citations-per-source) are not replicated; this is the analysis/Q&A core.
• Files API files auto-expire (~48h) on Google side; script also deletes them post-call.