Pathway: RAG Application Templates
/llm-app
Pathway AI Pipelines bundles ready-to-run RAG templates — question-answering, live document indexing, multimodal, private — on Pathway's own live-data engine instead of a stitched-together vector database and cache. Reach for it when retrieval must track a live file share or Kafka topic; skip it for a one-off prototype against a static PDF folder, where a plain script does less.
What are Pathway AI Pipelines?
Pathway AI Pipelines is a set of Docker-ready application templates built on the Pathway Live Data Framework, a Python library with a Rust engine inside. Each template wires document parsing, a built-in vector index (usearch) or hybrid full-text index (Tantivy), and an LLM call behind one HTTP API, connecting straight to Sharepoint, Drive, S3, Kafka, or PostgreSQL without a separate vector database.
Solving Real-time RAG Challenges
Most RAG tutorials assume a folder of PDFs you index once and query forever. Real-world RAG over Sharepoint, Google Drive, or a Kafka topic isn't like that — files get added, edited, and deleted constantly. A stale index quietly returns wrong answers. Pathway AI Pipelines treats data sync as a first-class part of the pipeline: the live-data engine watches the source and incrementally updates the vector and full-text indexes as changes happen, instead of you writing a re-indexing cron job yourself.
Core Capabilities for Live Data AI
- ✓Built-in vector index using usearch, plus a hybrid full-text index via Tantivy — no external vector database to run.
- ✓Live sync with the file system, Google Drive, Sharepoint, S3, Kafka, and PostgreSQL, so new, edited, or deleted documents update the index automatically.
- ✓Ships as Docker containers that expose an HTTP API; several templates add an optional Streamlit UI for quick demos.
- ✓Adaptive RAG template cuts token cost in RAG, per the README up to 4x, while trying to hold accuracy steady.
- ✓Multimodal parsing in the GPT-4o template pulls tables and charts out of PDFs, not just plain text.
- ✓Private RAG template runs fully local with Mistral and Ollama — no data leaves your machine.
- ✓Video RAG template uses TwelveLabs Pegasus and Marengo embeddings to make video content searchable.
Explore the Application Templates
- •Question-answering over a live document folder — the basic end-to-end RAG template for PDFs and DOCX files synced from Drive or Sharepoint.
- •Vector store / retriever backend for an existing frontend or a Langchain / Llamaindex app, via the Live Document Indexing template.
- •Financial-document analysis: the multimodal GPT-4o template extracts tables and charts, and the unstructured-to-SQL template turns report PDFs into queryable PostgreSQL tables.
- •Slide search across a library of PowerPoint and PDF decks with the Slides AI Search template.
- •Alerting when source documents change — the drive_alert example watches Google Drive and notifies you on updates.
Getting Started with a Template
There's no single top-level install command — each template under templates/ ships its own README with setup steps, since requirements (which LLM, which data connector) differ per template. The repo overall is Docker-friendly: templates run as containers, tested locally first, then deployed to GCP, AWS, Azure, Render, or on-premises. Exact commands per template aren't in the root README, so check the template folder you pick before assuming one docker run covers all of them.
Strengths
- ✓Swapping a data source or index type is a one-line pipeline change, per the README, instead of rewriting a service.
- ✓No separate vector database, cache, or API framework to stand up — the templates fold retrieval, indexing, and serving into one Pathway app.
- ✓Template variety covers more than generic Q&A: financial PDF-to-SQL, slide search, video RAG, and a fully local/private option with Mistral and Ollama.
- ✓MIT license leaves you free to fork a template and rebuild it around your own stack.
Understanding Pathway's Approach
- △No single quick-start command at the root — you have to open the template folder you want and read its own README before running anything.
- △GitHub lists Jupyter Notebook as the primary language, which fits experimentation better than a hardened production service out of the box.
- △The built-in vector index (usearch) and full-text index (Tantivy) are fixed choices — swapping to Pinecone or Weaviate means stepping outside the template's intended path.
- △Several templates depend on paid third-party APIs (GPT-4o, TwelveLabs) for their best results, so cost isn't zero even though the framework itself is MIT-licensed.
Comparing RAG Development Approaches
Common Questions
Pathway AI Pipelines connects to the local file system, Google Drive, Sharepoint, S3, Kafka, PostgreSQL, and real-time data APIs, syncing additions, edits, and deletions automatically.
Pathway AI Pipelines can run on-premises as well as on GCP, AWS, Azure, or Render — the templates ship as Docker containers, so the deployment target is your choice.
The templates work with GPT models by default, including GPT-4o for the multimodal template, and the Private RAG template swaps in Mistral running locally through Ollama for a fully offline setup.
Several Pathway AI Pipelines templates include an optional Streamlit UI for quick demos, but the core of each app is an HTTP API meant to be called from your own frontend.
Pathway Live Data Framework AI Pipelines is released under the MIT license.
Pathway builds a vector index in-memory using the usearch library, plus a hybrid full-text index via Tantivy, both updated live as source documents change rather than rebuilt in batch.
Who should try it — and who should skip
Reach for Pathway AI Pipelines if your RAG app has to stay current against a live file share, Sharepoint site, or Kafka stream and you don't want to hand-build the sync logic. Skip it if your documents are static and rarely change — a simpler LangChain or Llamaindex script against a one-time index gets you there with less to learn, and skip it too if you need a no-code UI rather than a Python codebase to fork.
Related repositories
Is llm-app worth your time?
ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of llm-app.