Beyond plain search: how PrismCI re-ranks to surface what matters

Ask a search box a hard question and it will hand you back a list — ten, twenty, fifty things that are sort of related — and leave the actual work to you. That's fine for a web search, where you expect to browse. It's a problem for competitive intelligence, where a rep has thirty seconds before a call and needs the answer, not a reading list.

Getting from "a pile of plausibly-related results" to "the right passage, first" is most of what makes PrismCI feel sharp instead of noisy. It comes down to two ideas: search more than one way, then re-rank.

Why plain ranking falls short

There are two common ways to search, and each fails in its own direction. Old-fashioned keyword search matches exact words — so it nails a product name or an acronym, but misses the point when the buyer phrases it differently. Modern "semantic" search matches meaning using AI embeddings — so it understands paraphrases, but it will also confidently return something that's merely in the same neighborhood, and it can fumble an exact term like a SKU or a competitor's feature name.

Rank by either one alone and the top of your list is noisy: near-misses that share the words but not the meaning, or fluff that shares the vibe but not the fact. For a CI answer that has to be right, that's not good enough.

Search more than one way, then fuse

PrismCI's retrieval runs on a model built exactly for this — BGE-M3, which brings several kinds of search together in one place: dense (meaning), sparse (exact terms), and multi-vector (matching at the level of individual words, not just the whole passage). Instead of betting on one method, PrismCI asks each and then fuses the results, so a passage rises to the top only when it's genuinely relevant by more than one measure. Agreement beats any single opinion.

One model doing all three keeps things simple — no separate stack for keywords, meaning, and fine-grained matching to run and reconcile. It also reads long documents whole (think competitor filings and contracts, not sliced-up fragments) and understands over a hundred languages, so a global competitive picture doesn't need a different search engine per region.

A result that wins on meaning, wins on keywords, and wins word-for-word is almost certainly the one you wanted. That agreement is the signal plain ranking throws away.

Then re-rank — the step that separates "related" from "the answer"

Fusion gets a strong shortlist. The last step makes it precise. A re-ranker — a cross-encoder — reads your actual question and each candidate passage together, at the same time, and scores how well that passage truly answers this question. It's slower than first-pass search, which is why you only run it on the shortlist, but it's dramatically more accurate: it's the difference between "contains your words" and "answers your question." On our own tests, turning the re-ranker on took Ask CI's top-result accuracy to effectively perfect.

There's a common misconception that a better AI model automatically means better answers. It doesn't. The model can only reason over what it's handed — so the search that decides what it sees matters more than the size of the model doing the writing. Get retrieval right and a modest model gives sharp, sourced answers; get it wrong and the best model in the world confidently answers from the wrong passage. PrismCI puts the effort where it actually moves accuracy: the retrieval layer.

The same idea, pointed at your feed

Scoring isn't only for answers — it's how PrismCI de-noises everything. The Intelligence Feed scores every competitor move for importance, so a real pricing or product shift rises and a cosmetic blog restyle sinks. Battlecard refresh scores which section a change actually belongs to before it suggests an edit. Same philosophy, everywhere: don't just gather signal, rank it, so people see what matters instead of everything.

Where it shows up

• Ask CI — the cited answer a rep or analyst gets is the re-ranked best passage, not a list to sift.
• Battlecards — drafting and refresh pull the right source, and match a change to the right section.
• RFP Assistant — each question is matched to the best answer in your library, with a confidence score.
• Intelligence Feed — every move scored for importance so the noise sinks and the shifts rise.
• Competitive Graph — the most relevant moves earn their place on the map instead of every pixel that changed.

Usable, accurate — and yours

The point of all this isn't the machinery; it's what it feels like to use. You get an answer instead of a search results page. You can trust it because it's the passage that actually fits, with the source attached. And because PrismCI runs this whole retrieval brain on your own hardware, that accuracy costs nothing per query and never leaves your network.