Small models, on your own machine
Run open GGUF models such as Qwen 2.5 and Llama 3.2 locally through llama.cpp. No API key, and your documents never leave the computer.
Logmark runs small open models on your own computer, builds assistants over your files, and shows the passage behind every answer. When the answer isn't in your documents, it tells you so instead of guessing.
Measured on a laptop with a 150,000-word PDF of India's Labour Codes and 12 test questions. A small test, shown so you can see how we measure.
Keep scrolling. The shape on screen follows your documents through each stage.
Drop in PDFs, Word files, spreadsheets, CSVs or plain text. Logmark extracts the text page by page and sheet by sheet, and everything stays on your machine.
Semantic chunking cuts on whole sentences into passages of about 150 to 225 words, with a little overlap so nothing falls between the cracks.
Hybrid search blends meaning-based embeddings with BM25 keyword matching, then a cross-encoder reranker rescores the best 30 candidates.
The model answers only from the numbered passages and cites them. If nothing relevant clears the bar, the model isn't called at all.
Logmark is the product; SLM Forge is the engine room behind it. Here's what ships in the current build.
Run open GGUF models such as Qwen 2.5 and Llama 3.2 locally through llama.cpp. No API key, and your documents never leave the computer.
Give an assistant a set of documents and instructions. It answers from them using retrieval, so the model's weights never change and there's no training run to wait for.
Every answer lists the files and pages it used. Answers without citations get a warning, and off-topic questions get a clear "not found".
Generate question-and-answer sets from your own documents, then score your assistants with AI-assisted (RLAIF) and human (RLHF) feedback before you rely on them.
Share an assistant you've built with a link, or drop a small chat widget onto a web page, like the one in the corner of this site.
When a job is too heavy for a laptop, you can choose to run it on Google Cloud compute. Day-to-day use stays on your own machine.
Retrieval runs in logmark-core, a Rust engine that keeps every document collection as a folder of plain files.
A manifest, a passages file and a vectors file per collection. Easy to back up, move or delete.
The manifest is written last, so a half-finished write is ignored. A lock stops two indexing runs from colliding.
The engine's server listens on 127.0.0.1 and only accepts requests from the app itself.
If the embedding model is missing, search falls back to keywords and says so instead of crashing.
# index a folder of HR policies > logmark-core index -c hr -p "C:\docs\policies" indexed 42 documents · 1,318 passages # ask a question, get cited passages back > logmark-core search -c hr -q "how many leave days" -k 5 [1] leave-policy.docx · p.3 score 0.91 [2] handbook-2026.pdf · p.17 score 0.84 [3] leave-policy.docx · p.4 score 0.78 # what's on disk corpora/hr/manifest.json corpora/hr/chunks.3.jsonl corpora/hr/vectors.3.f32
Leave, pay and benefits answered from your own handbook, with the page to check.
Ask across long codes and agreements and get the exact clause back.
Find the rule behind a process without searching a dozen PDFs.
Answer common how-to questions from your own runbooks.
Answers drawn only from approved material, with a clear "not found" otherwise.
Keep sensitive reading material on one machine and still search it by meaning.
Logmark is in alpha. Here's what works today and what doesn't exist yet, so you can decide whether it fits.
We're looking for a small number of teams with real documents and real questions to try Logmark, tell us what's missing, and shape what we build next.
Tell us about your team and what you're trying to solve. We review every application personally and reply within five working days.
By submitting you agree to our privacy policy. We will not share your details with third parties. Pilot places are limited.
Windows installer · alpha build 9.0.6
We're happy to have an informal conversation about whether Logmark fits your team. No commitment needed.