A locally-running AI service that executes intelligent workflows against a personal Obsidian vault. No data leaves the machine — everything runs on-device via a local LLM.
Two workflows are implemented: monthly_summary and semantic_search.
For over half a year I have been recording various aspects of my life in a git-synced Obsidian vault — daily notes, films watched, places visited, habits tracked. Over time it grew to the point where searching and summarising manually became impractical.
I wanted to bring AI capabilities to these notes, but the personal nature of the data ruled out any cloud service. LAIWM is the answer: a locally-running LLM pipeline that reads the vault directly and never sends data anywhere.
Machine: HP laptop — Intel Core Ultra with Intel Arc iGPU (17.92 GB shared VRAM).
The iGPU is the key hardware constraint that shaped every technical decision in this project.
| Setup | Result |
|---|---|
| LM Studio + Vulkan (Intel Arc) | Crashed mid-inference after ~4 batches |
| LM Studio + CPU only | Stable but ~2 hours per 30-note run |
| Ollama + IPEX-LLM (SYCL/oneAPI) | Stable + relatively fast — minutes per response |
IPEX-LLM provides a proper Intel GPU backend for Ollama via SYCL/oneAPI instead of Vulkan. Setup requires Miniforge, a dedicated llm conda environment, and a GPU-patched Ollama binary.
Model progression — each upgrade was justified by quality improvements:
| Model | Why |
|---|---|
qwen3:4b |
Initial choice — fit in RAM |
qwen3:8b |
Faster than the previous 4b despite being larger |
qwen3:14b |
Significant intelligence step up; stable on IPEX-LLM — current |
Embedding model: nomic-embed-text via Ollama.
Qwen3 has thinking mode enabled by default. Every prompt appends
/no_thinkto suppress it — without this, trivial responses take 10+ minutes.
- OS: Windows 11 + WSL2 (Ubuntu)
- Editor: VSCode with WSL extension
- AI assistant: Claude Code — used throughout development for implementation, debugging, and architecture decisions
- Python tooling:
uvfor dependency management
Reads all daily notes for a given month and produces a concise bullet-point summary of what actually mattered.
Pipeline:
collect month's .md notes
→ Pass 1: presummary per note (LLM × N)
each note condensed to 1-3 bullets independently
→ Pass 2: batch synthesis (LLM × batches)
condensed notes grouped and synthesized;
prior batch outputs passed forward to prevent repetition
→ Pass 3: final consolidation (LLM × 1)
all batches distilled to ≤25 key bullets
→ output file
Single-pass batching was tried first and abandoned — early notes lost detail as context grew. The three-pass approach keeps each note's contribution isolated before synthesis.
Example:
uv run manager.py --prompt "Summarize my 2026 march notes" --debug
[LAIWM 16:59:06] LAIWM — qwen3:14b @ http://172.x.x.x:11434
[LAIWM 16:59:06] Classifying request...
[DEBUG 16:59:20] Routing result: mode=monthly_summary confidence=100% params={'year': 2026, 'month': 3} missing=[]
[LAIWM 16:59:20] Mode: monthly_summary (confidence: 100%)
[DEBUG 16:59:20] Final params: {'year': 2026, 'month': 3, 'query': 'Summarize my 2026 march notes'}
[LAIWM 16:59:20] 31 notes → pre-summarizing...
[1/31] 2026-03-01 ... [DEBUG 16:59:58] Response (425 chars): - Went on a Kektura trip, ate sushi for breakfast ...
done
... (rest of the days were here)
[DEBUG 17:14:36] Synthesis: 31 day summaries → 2 batch(es), limit=10444 chars
[LAIWM 17:14:36] Synthesizing 2 batch(es)...
[1/2] 2026-03-01 → 2026-03-30 ... [DEBUG 17:16:04] ...
done
[2/2] 2026-03-31 → 2026-03-31 ... [DEBUG 17:17:06] ...
done
[LAIWM 17:17:06] Final consolidation...
[DEBUG 17:19:00] Response (1700 chars): - János worked on AI projects and set up local LLM environments for better GPU support - János registered an IBKR TBSZ account and invested in X - János helped with X’s diploma defense and s ...
[LAIWM 17:19:00] Output written to **/obsidian-llm/output/monthly_summary_2026_03.md
Answers a natural-language question by searching the entire vault. Combines vector similarity, keyword boosting, and question-focused summarization before synthesis.
Pipeline:
query
→ collect all .md files
→ embed files (nomic-embed-text, mtime-cached)
→ cosine score all notes against query vector
→ keep top 40% (drops vocabulary noise from unrelated notes)
→ LLM keyword extraction (content words only, no sentiment words)
→ keyword grep on top 40%
→ boosted score: vector × (1 + 0.1 × keyword_hit_count)
→ top 5 pool
→ presummary per note (LLM × 5) — each note summarized with the question in mind
→ synthesis on 5 focused summaries (LLM × 1)
→ output file
Pure vector search failed on notes with overlapping vocabulary. The hybrid pipeline — vector pre-filter → keyword boost → question-focused presummary — solved the recall precision problem.
Example:
uv run manager.py --prompt "In which town did I live while attending highschool?" --debug
[LAIWM 16:44:43] LAIWM — qwen3:14b @ http://172.x.x.x:11434
[LAIWM 16:44:43] Classifying request...
[DEBUG 16:45:14] Routing result: mode=semantic_search confidence=95% params={'query': 'In which town did I live while attending highschool?'} missing=[]
[LAIWM 16:45:14] Mode: semantic_search (confidence: 95%)
[DEBUG 16:45:14] Final params: {'query': 'In which town did I live while attending highschool?'}
[LAIWM 16:45:15] Stage 1 — scoring 473 notes...
[DEBUG 16:45:17] Kept top 40% (189/473 notes) by vector score
[LAIWM 16:45:17] Stage 2 — extracting keywords...
[DEBUG 16:45:24] Extracted keywords: ['town', 'highschool']
[DEBUG 16:45:24] Keyword matches: 26 files
[DEBUG 16:45:24] Synthesis pool (5):
[DEBUG 16:45:24] 0.677 Nyirbator Hungarian - English Two Language Elementary School.md | --- tags: - memo/chapters/education country: "[[Hungary]]" town: "[[Nyirbator]]" start: 2008-09 en
[DEBUG 16:45:24] 0.644 Eindhoven.md | --- tags: - memo/loc/towns/netherlands - lived-here type: town country: "[[Netherlands]]" region
[DEBUG 16:45:24] 0.626 Krudy Gyula High school.md | --- tags: - memo/chapters/education country: "[[Hungary]]" town: "[[Nyiregyhaza]]" start: 2016-09
[DEBUG 16:45:24] 0.618 Nyiregyhaza Apaczai Csere Janos Elementary School.md | --- tags: - memo/chapters/education country: "[[Hungary]]" town: "[[Nyiregyhaza]]" start: 2011-12
[DEBUG 16:45:24] 0.592 Nyircsaszari.md | --- tags: - memo/loc/towns/hungary - lived-here type: village country: "[[Hungary]]" region: Sza
[LAIWM 16:45:24] Stage 3 — pre-summarizing 5 notes with question in mind, then synthesizing...
[DEBUG 16:45:51] presummary Nyirbator Hungarian - English Two Language Elementary School.md: Rating: N/A (not applicable as no formal rating is given). Recommendation: N/A. Genre: Personal memoir.
[DEBUG 16:46:18] presummary Eindhoven.md: (ommitted)
[DEBUG 16:46:45] presummary Krudy Gyula High school.md: I lived in Nyiregyhaza while attending Krudy Gyula High School from 2016 to 2020.
[DEBUG 16:47:11] presummary Nyiregyhaza Apaczai Csere Janos Elementary School.md: Rating: 4/5. Recommendation: Highly recommended for its strong educational reputation and memorable experiences, though
[DEBUG 16:47:46] presummary Nyircsaszari.md: You lived in Nyírcsászári, a small village in Szabolcs-Szatmar-Bereg County, Hungary
[DEBUG 16:55:37] Answer (481 chars): You lived in **Nyiregyhaza** while attending high school, as explicitly stated in the summary about attending **Krudy Gyula High School** from 2016 to 2020. While the Eindhoven summary mentions living ...
# Semantic Search
**Query**: In which town did I live while attending highschool?
## Answer
You lived in **Nyiregyhaza** while attending high school, as explicitly stated in the summary about attending **Krudy Gyula High School** from 2016 to 2020.
## Sources
(ommitted)
Requires Miniforge with the llm conda environment configured for IPEX-LLM. Open a Miniforge terminal:
conda activate llm
set OLLAMA_NUM_GPU=999
set ZES_ENABLE_SYSMAN=1
set SYCL_CACHE_PERSISTENT=1
set no_proxy=localhost,127.0.0.1
set OLLAMA_HOST=0.0.0.0
ollama serve
OLLAMA_HOST=0.0.0.0is required — without it Ollama binds only to127.0.0.1and is unreachable from WSL2.
Pull the required models (one-time):
ollama pull qwen3:14b
ollama pull nomic-embed-text# Clone and install
git clone https://github.com/GJanos/obsidian-llm
cd obsidian-llm
uv sync
# Set vault path (add to ~/.bashrc to make permanent)
export OBSIDIAN_LLM_VAULT_PATH="/path/to/your/obsidian/vault"
# Run
uv run manager.py --prompt "your request here"
# Run with debug output
uv run manager.py --prompt "your request here" --debug