Contextual Priming: Shaping Trajectories In AI/LLM Search Spaces
Published by Daniel A. Dahlen on
“Diamond.”
In a previous post here I mentioned how SEOs were the attention mechanism before the modern day transformer attention mechanism – guiding and shaping relevance in traditional search spaces.
In old school/traditional search, this meant accumulating enough (relevant) context on and around a web page, website, brand/entity, et. al. to an “accumulation point” where that particular page/site/brand/entity became included in the most relevant candidate pool in any particular search or category.
If you were the most relevant – or had accumulated enough context, that is – you were positioned higher on a search results page. Better positioning meant increasing the probability of a user choosing that page/site/brand/entity – something the industry calls “clickthrough rate” (CTR).
Stepping back, CTRs in search closely resemble transition probabilities we know in graph theory and walks – if the aggregate profile [ query + user + search context ] represents the starting node, then the list of relevant candidates represent different paths through the search space a user can take or click or tap – each with their associated transition probability; the odds that a user passes through those candidate nodes on their “information walk”, of sorts.
We often forget that the search engine itself (and the user) are part of the graph — but I digress.
Over time as those walks are exposed and repeated, the stronger the association becomes between candidate pages and a particular aggregate profile (query/users/search context). You can imagine that this repetition can improve a search engine’s confidence and ability to retrieve those candidates in those moments – in terms of both accuracy and speed.
At this point, you may be wondering what “diamond” has to do with any of this.
“Diamond” was my nickname growing up (the origin of which I’ll share some day), but we all know how strongly nicknames stick over time.
Through conversations – written or spoken – nickname usage repeated by relevant groups draw it closer and closer to a person’s real name. At first that nickname could represent many different meanings – “diamond”, of course has different “paths” or directions of meanings, depending on the user’s particular state or association.
As those particular users repeat that nickname when referencing that person, it improves those user’s ability to remember (or “retrieve”) the real person that nickname is associated with.
Those users may even correct people when they mention that person’s real name:
“Dan? Oh… Diamond, yeah he’s over there.”
The common thread between these two examples – improving accuracy or speed of retrieval through repeated exposure is something called priming. (There are different kinds of priming, but we’ll stay general here in this particular context.)
Extending Priming Intuition Through Graph Theory
Touched on briefly above, priming can be thought of as shaping transition probabilities of a walk through a graph (check out my primer on graph theory, Markov chains and eigenvectors if you want catch up on transition probabilities).
Repeatedly exposing two (or more) words together – for certain users – (“diamond” and “Dan”) increases the probabilities of someone’s next thought “step” going from “diamond” to “Dan” (and, in some cases, vice-versa).
As context accumulates – through repeated exposure – “Dan” will almost always be chosen as the next “step” when seeing/encountering “diamond” (approaching 100% probability) – in relevant situations. There won’t be a hesitation – just an immediate response/memory.
When viewed as a compressed memory – that accumulated context and retrieval event can be viewed as a Markov chain – where the next state (or node) only requires information about the immediately preceding state or step (“diamond” always transitions to “Dan” through a compressed memory of “diamond” for a particular user).
Swap “nickname” for brand name and “real name” as a generic solution or product/service, you’ll have some marketing magic (search + marketing magic, really).
Priming Changes For Different Users & Projection
For different users the word “diamond” will step into the traditional/consensus meaning – the crystal/mineral diamond, of course, or some other personally associated/consensus meaning they may have.
Priming of a certain words changes between different groups of users.
Extrapolating that to our graph intuition, different users will map to different transition probabilities and have different transition tendencies, essentially, when encountering different words.
Projection – something I wrote about a few weeks ago – or the mapping of some high dimensional space down to some lower dimensional subspace, relates nicely to priming in this sense.
Each user projects their own primed states onto words (the shadows in ghosts) — collapsing into different graphs of meanings or associations depending on how their personal transition probabilities are shaped through previous priming whenever they encounter that word.
Priming In AI/LLM Search (Contextual Priming)
Extending word priming to question-answer associations (or prompt-response), this intuition can be generalized to larger and larger “decision units” (deciding which branch or meaning path/walk to take).
Instead of a users’ memory (that projects their own primed meanings onto words) you can imagine an AI/LLM search space as one collective associative memory (Hilbert) space, that comes with its own priming tendencies for different prompts/queries.
I see each response from an AI/LLM search surface as a dynamic walk through this space, governed by an evolving transition matrix at each step that has hints – or ghosts – of the Markovian nature of our traditional web graph navigation (users transitioning from one web page to the next through links).
Generalizing, there’s a seeker of information (user/query/prompt/context), a determination of relevance (transition matrix + attention or link) and a payload of information (the response/word at each step – or set of web pages/sites/brands in our search analogy).
What we see within that governing, evolving transition matrix are the accumulated/residual priming effects of each model, shaping the paths or walks through the space for each prompt (which is why, I believe, each model “feels” a little different – like speaking with two different people with different primed states).
The trajectory of each response “walk” or “path” reveals the transition tendencies between words, concepts, entities, brands, etc. – and those primed graphs of the underlying space for each model.
From a technical perspective they might call these training effects or parameter tuning — but if you step back a bit, one can interpret those as “contextual priming” – improving the speed and accuracy of responses.
The Million (Billion?) Dollar Question: Can We Shape Response/Conversational Trajectories?
I’ve never been much for technical/granular gimmicks or “tricks” — they just temporarily distort priming/transition tendencies or the underlying graph/space for a time before returning to more stable (eigen) states, so you won’t find any of those here.
Knowing your users and accumulating context onto (and within) a high fidelity representation for a particular topic/vertical.
One gives you a stable target/identity (high fidelity representation), the other gives you the handle/controls for priming (accumulated, categorical/vertical context) the paths or walks through these spaces. The user set constrains and personalizes the available context, limiting it to what’s most relevant to them.
No tricks, no manipulation (proper search optimization has never been about manipulation) — just clear, consistent information about who you are and accumulating enough context that response trajectories that contain (or start with) those contextual elements have no choice but to step through your information/location in the space in order to produce a response – for a particular set of users.
The more dense the space (a region in the model with many relevant, strong candidates for some input – or operator – space), the more context (or more time) you may need to accumulate to pull that priming in your direction, but ultimately the machinery is the same as it’s always been in search.
Up Next
Knowing how many different paths/walks (and the operator space that produces those walks) you’ll need to account for to properly approach measurement (from the outside looking in at least) becomes the next logical item on our agenda, then.
The contextuality effect present in AI/LLM search spaces throws us some curveballs, but luckily the non-classical world offers us some tools to help harness and predict things – and improve the odds of better measurement outcomes.
As for this post, I’ll leave you with this question:
The next time you see or encounter the word “diamond” – what will you think of?
