2023 - present
Knowle
The Agent Toolkit Behind Our Clients' AI Features
Our roleProduct ownership, toolkit architecture, LLM integration, client deployments
3 daysfirst client POC on the toolkit
3 productsclient AI features shipped
3 channelsSlack, web, WhatsApp
MITopen source, self-hosted
Knowle is LambdaWorks' own toolkit for building AI agents. It started in 2023 as a Slack bot answering questions from our internal documentation. In 2025 we refactored it into a library our teams pull into client products, and it keeps growing with every deployment - connectors, retrieval, agent tools, and delivery channels already solved, so a working agent takes days, not months. Open source, MIT licensed.
The problem with building agents from scratch
Every AI feature restarts from zeroEach new agent needs the same plumbing: document ingestion, chunking, embeddings, vector search, tool calling, streaming, source citation, feedback capture. Teams rebuild it per project and pay for it three times.
Agents have to reach people where they workAn agent behind a chat box on a marketing site gets used once. The same agent in Slack, in WhatsApp, or embedded in the product gets used daily. Each channel is its own integration problem.
Model and infrastructure choices keep movingHard-wiring one LLM provider or one vector database turns every re-evaluation into a rewrite. A year of model releases makes that a recurring cost.
What we built
Knowledge connectorsGoogle Drive, GitHub, and FTP sources sync into the index. A record manager keeps re-indexing incremental instead of rebuilding the whole collection on every push.
Swappable retrievalVector search across Qdrant, Milvus, or PostgreSQL/pgvector. The provider is a configuration value, not an architectural commitment.
Agent runtimeLangChain-based reasoning with tool calling. Responses stream token by token, then return the source documents they were built from, with feedback captured per answer.
Agent toolsIntegrations the agent can call: Clockify, Google APIs, and company data services. New tools plug into the same interface.
ChannelsSlack bot, web app or embedded widget, and WhatsApp Business API. BullMQ handles scheduled and long-running jobs.
Multi-modelOpenAI, Groq, and Mistral behind one factory. Model and temperature are set per deployment.
Where it runs
Knowle is not a demo. It ships inside client products.
SuperPhoneFirst version of an AI agent letting users operate the platform through conversation instead of navigating the UI.
ExportOnlineAI customer support assistant on the vehicle-export customs platform, answering from the client's own documentation.
EveryBiteProof of concept built on Knowle in three days, from a standing start to a working demo.
The stack
TypeScript end to end. NestJS on the server, React with MobX on the client. LangChain for retrieval and agent reasoning. Qdrant, Milvus, or PostgreSQL/pgvector for embeddings. OpenAI, Groq, or Mistral for inference. BullMQ for job processing, Knex and Objection.js over PostgreSQL, Auth0 for authentication, Slack Bolt for the Slack client.