A short glossary
What we do, in plain words
Seven ideas behind Flowlexi's retrieval, workflow, and cloud systems — explained for people who run businesses, not servers. Each one links to where it lives in the stack.
Semantic search
Keyword search looks for matching words. Semantic search can find related passages even when they use different wording: a search for “contract cancellation deadlines” may surface a clause saying “either party may terminate within 30 days”.
It works because text is compared by meaning, not spelling. For a business this changes what a knowledge base is worth: people find answers phrased differently from their question — which is most answers.
PaveDB is a search engine built this way — pavedb.org Back to top ↑
Embeddings
An embedding model reads a piece of text and places it as a point on a very large map, where distance means similarity of meaning. “Refund policy” and “money-back rules” land close together; “refund policy” and “parking rules” land far apart.
Semantic search is then just: place the question on the same map, look at what sits nearby. The model and its version influence the results, alongside how documents are split, filtered, and ranked.
An inspectable vector database
A vector database stores those points and searches for nearby items. To review a result, you also need to see the original passage and where it came from.
An inspectable system connects results to their sources and keeps the query and ranked results in a log. Replay reruns that query against the current collection. Compare the records to see how results change as documents are added or updated, before using them to support a decision.
Book · public draft coming soon — inspectableretrieval.com Back to top ↑
RAG — retrieval-augmented generation
Language models are fluent but they answer from memory. RAG makes the model look things up first: retrieve the relevant documents, then write the answer from them, with citations.
The quality ceiling of a RAG system is its retrieval. If you cannot audit what was retrieved and why, you cannot explain the answer — which is why we start from the search side.
Portable, declarative AI flows
Most AI automations are glue code: logic buried in scripts only their author can follow. A declarative flow states the steps in a readable file — retrieve, decide, act — like a recipe rather than a trail of pans.
Readable files can be reviewed, versioned, and moved between environments. The automation stops being a person’s secret and becomes a company asset.
Self-hosted models
The models that read and write text — encoders and decoders — can run on hardware you control instead of a third-party API. You choose where inference runs and which model version to use. Data stays in your environment when the rest of the application is also configured that way; you still need to budget for hardware and operations.
The trade-off is honest: you operate the machine. That is exactly the part we sell as a service when you’d rather not.
Managed instances, your keys — cloud.flowlexi.com Back to top ↑
The cloud console
The console brings managed retrieval, tenants, API keys, collections, and billing into one interface. Your team uses the service while Flowlexi operates the infrastructure.
PaveDB is open source. Run the same engine inside your Python application, as a service on your infrastructure, or managed by Flowlexi. Collection export and import let you move your retrieval data between deployments.
Start with managed PaveDB — cloud.flowlexi.com Back to top ↑
Want to see inspectable retrieval running? Start with a managed PaveDB instance.
Start free with managed PaveDB