The Challenge
Litify is deep, versioned, and specific. Consultants, analysts, and developers ask detailed product questions every day: how a Docrio feature behaves, which fields a standard object ships with, what changed in a release, whether a requirement can be met out of the box. Getting those answers right matters, because a wrong one does not stay a wrong answer. It becomes a requirement, a design decision, or a commitment to a client.
General-purpose AI assistants are fluent but unreliable on this kind of detail. They will describe a setting that does not exist or a behavior from the wrong release with complete confidence. Turning on open web search adds new problems: misinformation, prompt injection, and no control over what the model reads. The fallback was manual work: searching the Litify community knowledge base article by article, or digging through tens of thousands of raw metadata files to learn what the base package actually contains.
BRIO needed an assistant that could:
- Answer Litify product questions from trusted sources only, never from general knowledge.
- Cite every answer so a person can verify it in seconds.
- Say clearly when the sources do not contain an answer, instead of guessing.
- Account for the specific Litify versions installed in a client’s org.
- Be shared safely by the whole team and by BRIO’s other AI agents, at a sustainable cost.
The Approach
Rather than hoping a general model would get Litify right, BRIO built a retrieval-augmented generation (RAG) system around three trusted sources, then wrapped it in an agent with explicit guardrails. The sources were chosen to complement each other: the Litify knowledge base for how the product is documented, a complete dissection of the Litify base package for how it is actually built, and BRIO’s own internal Litify channel for the fixes and workarounds the team has learned on real projects.
Governance came first, not last. Before anyone relied on an answer, the team defined what the agent may read, what it must cite, how it labels confidence, and what it must refuse to guess. Every stage of the pipeline was verified before the next was built.
The Solution
A Curated Knowledge Index: BRIO indexed 655 articles from the Litify documentation available to it as a Litify partner, across product guides, release documentation, troubleshooting, Salesforce, and integrations. Cleaning preserves tables, heading structure, and list numbering, and each article is tagged with the release it describes.
AI-Assisted Chunking With Rules: Claude splits each article into self-contained passages under strict rules: never break a table or a numbered list, keep every passage understandable on its own, and carry the release context forward. Automated checks confirmed that no content was lost or altered along the way.
Context-Aware Embeddings: The resulting 3,978 passages are embedded with a contextual model that understands each passage in light of the article around it, with release tags included so version-specific content is found for version-specific questions. Search is exact and returns results in under a second, without additional database infrastructure the current corpus does not need.
Grounded Answers Only: For each question, the agent retrieves the most relevant passages and answers strictly from them. It may cite only sources that appear in what it retrieved, and when nothing relevant is found it says so. Follow-up questions are understood in context, so a vague “what about the Salesforce version?” works without repeating the whole question.
The As-Built Package, Not Just the Docs: BRIO extracted a complete metadata map of an untouched Litify org, covering 385 objects, 3,788 fields, 64 flows, and hundreds of reports, into a searchable, version-labeled reference. The agent uses it for exact API names, field defaults, and flow logic, and it doubles as a test for customization: anything in a client’s org that is not in the baseline is, by definition, custom.
BRIO’s Field Knowledge: Alongside official documentation and as-built metadata, the agent searches BRIO’s internal Litify channel for fixes the team has already worked out, and labels them clearly as BRIO workarounds rather than product documentation.
Version Awareness Built In: When an answer will be used in a live org, the agent first confirms which Litify package versions are installed and answers for those versions. It states the newest release it can actually read, flags articles that carry no version marker, and treats a version supplied in a question as a claim to check rather than a fact.
One Expert, Available to Every Agent: The knowledge expert is delivered as a single tool that any BRIO agent can consult, from the business analyst agent confirming what Litify supports out of the box before a spec reaches a client, to the data and file migration agents checking data model and document behavior. People can use it directly through a BRIO-branded chat that signs in with company credentials.
Governance and Accuracy
Governed by Design: The agent is read-only. It answers only from its three sources, never from general knowledge, and it describes the product rather than any client’s private configuration. The metadata reference contains no client records and is checked for personal information.
Every Claim Is Labeled: Each answer cites its source and says what kind of source it is: official documentation, as-built metadata, or a BRIO team workaround. The agent distinguishes what an article states from what it infers, and it never presents an inference as a citation, because an inference that looks verified is the hardest error for a reader to catch.
“Not Found” Is a Feature: A clear “the sources don’t cover this” is treated as a better answer than a plausible guess. A known gap becomes something to test in the org; a confident guess becomes a promise nobody can keep.
Verified at Every Stage: Article counts were checked against the source, every cleaned article was compared to its original, passage splits were checked automatically and reviewed by hand, off-topic questions were confirmed to return “not found,” and the metadata reference passes its own evaluation suite.
Cost and Efficiency
Right-Sized Models: Each step uses the least expensive model that does the job well. A capable mid-tier model handles chunking and answering rather than the largest available, keeping the cost of each question small.
Lean Infrastructure: The service runs on a modest cloud tier, and search runs without a separate database layer until the corpus is large enough to need one. The design choice is reversible, so the system can grow without being rebuilt.
Minutes Become Seconds: Questions that once meant searching the knowledge base article by article, or combing through raw package metadata, now return a cited answer in seconds.
Technical foundation
- Claude for AI-assisted chunking and grounded answer generation
- Contextual embeddings with release-aware tagging
- Exact similarity search with sub-second response
- Model Context Protocol (MCP) tool shared across BRIO’s AI agents
- Version-labeled Litify base package metadata reference (JSON and searchable database)
- Cloud-hosted service with a company sign-in chat interface
The Impact
The knowledge expert has become shared infrastructure for BRIO’s Litify practice:
- Cited, version-aware answers to Litify product questions in seconds instead of manual searching.
- Out-of-the-box versus custom determinations grounded in the actual base package.
- Used by consultants for package upgrade planning and by BRIO’s agents during requirements and build.
- Governance, accuracy, and cost designed in from the start rather than added later.
Next, the team plans to index BRIO’s own documented fixes directly and extend the expert into common administrative workflows.
There is no shortage of people explaining RAG. What makes this different is that BRIO built it, governs it, and uses it on real client work, and the most important thing it learned to say is “I don’t know.”
