Alice uses a knowledge base to learn about a seller from documents, websites, recordings, and other source material.
2
The ingestion pipeline converts resources to markdown, chunks the content, stores embeddings in Pinecone, and retrieves context through a deep research agent.
3
The team recommends getting a workable system into production before benchmarking vendors and tuning cost or quality.
Summary
Sherwood Callaway explains how 11x built the knowledge base behind Alice, its AI sales development representative. The old product required customers to manually enter products, pain points, value propositions, and other details into a library before they could run campaigns. That created onboarding friction and often gave the model either too little or too much context. The new system lets customers upload documents, websites, audio, and video. Vendors parse these resources into markdown, the system splits the markdown into structured chunks, and Pinecone stores the embedded content for similarity search. Alice uses a deep research agent to plan retrieval, gather relevant context, and answer questions that support email generation. A 3D view of the vector space and a Q&A interface let customers inspect what Alice knows. The team is still working on hallucination tracking, parser evaluation, hybrid retrieval, and pipeline cost.
Alice needs knowledge about both the seller and the lead
An SDR sources leads, contacts them across channels, and books meetings. For Alice to write useful emails, she needs information about the seller's products, services, case studies, pain points, value propositions, and ideal customer profile. She also needs information about each lead, including their role, responsibilities, concerns, previous solutions, and company. This talk focuses on the seller side. The seller provides business context, and Alice uses it to personalize messages for individual leads.
The manual library created onboarding friction and weak context selection
The old product asked customers to define every product and service in a library, along with detailed pain points, solutions, and value propositions. Customers could not run campaigns until they completed that setup. Campaign creation then required users to select which offers Alice could use for a particular lead. Choosing too few offers could make the message irrelevant. Choosing too many filled the context with unrelated material and made Alice less effective when writing the email.
The knowledge base lets Alice pull in seller information from source material
The team changed the workflow so customers can upload their existing material instead of manually entering all business context. The knowledge base is a central repository for seller information, similar to onboarding a human SDR by giving them documents and letting them learn over weeks or months. The supported resource groups are documents and images, websites, and media such as audio and video. Alice can reference that material when generating messages.
The ingestion architecture is asynchronous and resource-specific
A user uploads a resource in the client, where it is saved to an S3 bucket and sent to the backend. The backend creates database records and starts jobs based on the resource type and selected vendor. Vendors parse the resources asynchronously and send webhooks back to the system. The parsed artifact is stored in the local database, then upserted and embedded in Pinecone. The interface updates after storage, and Alice later queries the vector database for the information.
Parsing turns PDFs, websites, and media into structured markdown
Parsing converts non-text resources into text because language models work with text. The team chose markdown because it preserves useful structure and formatting, such as titles and paragraphs. They decided to use specialist vendors rather than build parsers internally, since the system had several resource types and many file formats. LlamaParse handled documents and images, Firecrawl handled websites, and Cloudglue handled audio and video, including information extracted from the video itself rather than only its transcript.
Chunking preserves meaning while keeping retrieved records manageable
After parsing, the system has long markdown documents that must be divided into semantic units for embedding and retrieval. Possible strategies include splitting on tokens, sentences, markdown headers, or using an LLM to choose boundaries. 11x uses a waterfall combination: it splits on markdown headers, then sentences, then tokens. This keeps structural meaning intact, avoids overly long chunks that dilute the document's meaning, and produces chunks that can also be displayed cleanly in the product interface.
Pinecone supports the similarity search the system needs
Retrieval-augmented generation can use many storage systems, including graph, document, relational, and key-value databases or object storage. 11x chose a vector database because Alice needs similarity search. The team selected Pinecone because it was cloud hosted, easy to start with, had useful SDKs and guides, and included embedding models. Customer support also helped the team compare vector database and graph database options.
Alice uses a deep research agent to plan retrieval
The retrieval system evolved from traditional RAG, where retrieved context is added to an LLM prompt, to agentic RAG, where retrieval tools are called inside an agent flow. 11x built what Sherwood describes as a deep research agent using Leta. The system passes in lead information, creates a plan with one or more retrieval steps, calls the tools, summarizes the results, and produces a clean question-and-answer response for the messaging process.
Visualization gives customers a way to inspect Alice's knowledge
Customers want evidence that Alice knows their products and will not invent case studies, testimonials, or pain points. 11x projects the vectors from Pinecone into three-dimensional space and renders them as nodes. Users can click a node to view its associated chunk. The product also provides an upload flow, a Q&A agent for interrogating the knowledge base, and a campaign view that shows retrieved chunks behind the questions used in the messaging flow.
"The first is that she needs to know the seller, the products, the services, the case studies, the pain points, the value props, the ICP."02:38
Who should watch
You are building a knowledge ingestion system for an AI agent and need practical choices for parsing, chunking, storage, and retrieval.
Your product asks customers to manually teach an agent about their business, and onboarding is becoming a source of friction.
You are comparing RAG vendors and need a production-first approach to evaluating quality, cost, and retrieval design.
