Langflow 1.10 just added IBM Db2 as a native vector store.
I haven't used Langflow yet. If you haven't either, you're not alone. I'm right there with you.
I've watched a colleague Gregory Petrossian build AI workflows with it.
Today I read the docs, watched the demos, and wrote down the 16 questions that came up.
I looked up their answers in the official Langflow documentation and ran them past Eric Hare from the Langflow team to confirm.
If you're new to Langflow: it's an open-source, Python-based framework for building AI applications. Drag components onto a canvas, wire them together, get a working RAG pipeline or agent.
The new Db2 Vector Store component plugs straight in. No glue code. Drag, connect, run.
The Q&A below covers code requirements, runtime, document ingestion, chunking, embeddings, LLM providers, prompts, the Db2 component itself, deployment, monitoring, and guardrails.
Background reading:
1. The basics
Q1. What is Langflow?
Langflow is an open-source, Python-based, customizable framework for building AI applications. The visual editor lets you drag components onto a canvas and wire them together into working pipelines. Common use cases include RAG, agents, ingestion workflows, and chatbots.
Q2. What is a "flow" in Langflow?
A flow is one pipeline you build on the canvas. Boxes wired together, saved as one unit, that you can run end to end. For example, File → Split Text → Embeddings → IBM Db2 Vector Store is an ingestion flow. Each flow has its own ID and REST endpoint. One project holds many flows, and one flow can call another.
Q3. Do I have to write code?
No. The visual canvas covers the full pipeline. If you want more, every component lets you open its Python source and edit it, and you can create new components from scratch. So it's no-code first, with a Python escape hatch when you need it.
Q4. CPU or GPU?
CPU. Langflow orchestrates, models run elsewhere. GPU only matters if you host an LLM locally.
Q5. Templates or build from scratch?
Both. Langflow ships Starter Projects (basic RAG, agent with tools, structured data agent, and more) that you can clone and modify. You can also wire components from scratch. In 1.10, the Assistant generates a full flow from one sentence.
2. Loading documents
Q6. What document formats, and can it pull from a URL?
PDF, DOCX, TXT, HTML, CSV, Excel, Parquet, JSON, TSV. A URL component pulls straight from the web.
Q7. Does it extract text from PDFs? Does it use Docling?
Yes to both. The lfx-docling bundle turns PDFs and Office docs into clean Markdown.
3. Chunking
Q8. Does Langflow chunk for me, and what strategies are available?
Yes. Drop in the Text Splitter component. Defaults work out of the box, no chunking expertise needed. Strategies in a dropdown: character, recursive character, token-based, and language-aware (Python, Markdown).
4. Embeddings
Q9. Cloud embeddings, local, or both?
Both, same slot on the canvas. Cloud: watsonx, OpenAI, Cohere, Mistral. Local: Ollama, LM Studio, vLLM, Hugging Face.
Q10. Text only, or multimodal?
Text. No first-class multimodal embedding component yet.
5. LLMs and prompts
Q11. Which LLM and embedding providers ship out of the box?
About 60 bundles. LLMs: OpenAI, Anthropic, IBM watsonx.ai, Google Vertex, Mistral, Cohere, Groq, Ollama, vLLM, Azure, NVIDIA. Most also offer embeddings.
Q12. Predefined prompts only, or can I write my own?
Write your own. The Prompt component is a free-text editor with {variable} placeholders that bind to upstream inputs. No fixed template library you have to use.
6. The IBM Db2 Vector Store
Q13. Is using the Db2 Vector Store really no-code?
Yes. uv pip install langflow pulls in the lfx-ibm bundle and the ibm-db driver automatically. After that, drop the IBM Db2 Vector Store component on the canvas, fill in hostname, port, database, credentials, table name, and an Embedding component. Drag, connect, run.
7. Using your flows from other apps
Q14. How do other apps call my flows?
Every flow gets a REST endpoint: POST /api/v1/run/{flow_id} with an API key. Webhook and OpenAI-compatible endpoints also work. Or publish the flow as an MCP server.
8. Monitoring and safety
Q15. Can I monitor what's happening inside the pipeline?
Yes. Set environment variables for LangSmith, Langfuse, or LangWatch and Langflow streams traces automatically. You get per-component latency, token counts, the exact prompts sent to the LLM, and cost estimates. No code changes to your flow.
Q16. Guardrails and hallucination detection?
The Guardrails component blocks PII, secrets, jailbreaks, prompt injection, offensive content, and malicious code, plus custom rules you write. Policies (Beta) guards agent tool calls. For hallucinations, the Cleanlab bundle scores response trustworthiness in real time.
What's next
This is my introduction to Langflow and the new Db2 Vector Store component. I haven't used it yet.
Next on my plan: actually build something with it, and write a follow-up post about what I learn.
How about you? Are you going to try it out?
If you have questions, drop them in the comments and I'll dig in.
Further reading