Context Graph for CPG

David: What does “context” mean for data platforms?

Ophir: Ever since we started tackling the problem of messy, fractured product data, we found that domain knowledge is the key to making good decisions.

Context in AI goes beyond recording state data and decision traces. It must include the basis of the decision. For specific verticals — like CPG — the reasoning is usually embedded in industry-specific domain knowledge.  

David: What does this mean in real life?

Ophir: A key aspect of our data platform is that industry knowledge is embedded in our logic. But the more expertise that we added, it became harder to manage, especially for performing QA.

This problem led us to add human-readable “context” right alongside the data results. This approach was powerful and efficient. It enabled agents, analysts, and developers to trace the chain of decisions for audit or validation.

Context is embedded vertically in our stack. Throughout our schema and processes, we can observe the entire "sausage-making" process, with natural language traces that lead clearly to the final outcomes.

David: Can you give some examples of domain knowledge about CPG products?

Ada: Sure, for example: How are units of measure correctly expressed in food versus non-food products? We see unexpected mutations all the time, such as made-up or foreign acronyms, even from “sources of truth” such as brand syndication and major retailers.

John: Another example: FDA regulations can be used to validate Nutrition Facts data. The FDA has strict requirements, which are codified in the Federal Register and have the force of law, but are often not applied on product description pages.

David: We see all the time how FDA regulations about food labeling, as expressed in the FD&C Act, are often violated on retail websites.

How do retailers get away with it? Because unlike food brands, grocers are not subject to federal oversight of product description pages.

Ophir: These are good examples of how we use domain knowledge to make decisions about normalizing, fixing, and enriching attribute values. While we don’t know about travel and automobiles, we do know CPG.

David: How does recording context help?

Ophir: Context is necessary to describe why decisions were made.  It must be easily observable, auditable, and stored as first-class data.

David: Aren’t some data sources always reliable, so context about decision-making isn’t needed?

Ophir: Among the 47 sources we currently ingest, all have significant gaps and inaccuracies.  

Ada: Some data sources have excellent quality, but only cover a subset of brands. Others, like syndicators, don’t have private label products. And long-tail SKUs are specific to each source, or not present at all.

John: Yes, and key attributes are often missing, such as images or net weight.

David: Sounds like a mess. How do we deal with it?

Ophir: We expect data chaos. Gaps and errors are the problems we solve for.

David: Aren't LLMs a reliable source of product data?

Ophir: Looking at generic AI results, they do a great job at synthesizing and generating data. But general LLMs don’t have vertical knowledge to classify with accuracy, correct errors, derive missing values, or adjudicate conflicts.  

Ada: The answers from LLMs are only as good as the information they gather. We often see inaccurate results from general AI search and tools. Garbage in, garbage out still applies in the age of AI.

David: Is the answer to throw an army of human experts at the problems?

John: Human experts do have the knowledge, but an army of them is expensive and hard to manage. However, their knowledge can be embedded in vertical AI to create quality data at scale.

David: When did we start using context?

Ophir: We architected context from the beginning. We just didn’t call it a “Context Graph” because that term didn’t exist. Handling context as first-class data enables our systems and people to do their jobs better. We embed our expertise to improve quality and frequency, then capture how that expertise was used to create those better outcomes.

Ada: A simple example is knowing that cocktail mixers like tonic are non-alcoholic, even if retailers or brands list them in an alcohol category.  

John: LLMs often get this classification wrong too. Humans like Ada never do 🙂.

David: These CPG nuances matter in lots of use cases, such as PDPs, digital analytics, nutrition and even price optimization.  

Ophir: As our contexts grow, our learning compounds. The more we add to the graph, the more we are able to do better. Compare this context approach with hard-coded rules for addressing endless edge cases. That becomes unwieldy and doesn’t scale.

David: When does the Context Graph come into play?

Ophir: In our view, a Context Graph is systems of agents that enable autonomous decision-making across workflows. This full AI autonomy — when AI makes decisions across workflows — has not yet arrived. We see the benefits and the paths to get there, and we are working towards them.

David: Where are we on the journey to a Context Graph?

Ophir: The first requirement is having an AI Native system, with schemas and processes that are designed for context, not just state and observability. We re-architected our systems to be fully AI Native in 2024.

Second, we treat context data as “first class.” This means the "Why" data is of equal importance to all other data types.

Third, our agents already utilize context data to decide, store, observe, and audit decisions in discrete workflows.

We believe that the full concept of the Context Graph will be realized when systems of agents — not just specific services — can exercise autonomy across workflows with provable superiority.

David: Why not sacrifice quality to be more efficient?

Ada: Our value proposition is strongly rooted in quality data.  

John: We aggressively lean into AI, but we won’t “ship it” if the results are not better or cause harm.  

Ophir: We see a clear path to the Context Graph. We believe it describes how data businesses will scale in the age of AI, with quality and efficiency, grounded in domain knowledge.

David: Thank you, Ophir, John, & Ada. Glad to be on this journey with you!

For our first mention of the Context Graph, see our recent blog post The CPG Showdown: Love vs Strength.
‍
For more on how context graphs are being discussed across the tech and venture communities, see What are context graphs by Simple.AI and Context graphs one month in by Foundation Capital.
‍
For the original Context Graph idea, see AI's trillion-dollar opportunity: Context graphs, by Foundation Capital.‍

Thank you for following our progress. Send me a note with thoughts or questions. I’d love to hear how we can support your work.

Warm Regards,

David Goodtree

Founder and CEO, Foodgraph