Market Research Knowledge Base: Answers That Cite the Page

A market research knowledge base fails at retrieval, not at knowledge. What the measured numbers show, and why every citation has to come from a lookup.

A category director asks whether anything in the last five years explains why shoppers aged 25 to 34 walked away from the large pack. The answer exists. It sits in a 2023 usage study, in two focus groups from 2024, and in an open ended question coded in 2021. Nobody remembers which folder. That is the case for a market research knowledge base, and the question most archives cannot answer.

Key Takeaways

• Google Research reports that frontier models hold 95 to 98 percent of the facts tested and fail to retrieve 26 to 34 percent of them when answering. The gap is retrieval, not knowledge.
• In the same work, letting citations come from the model's memory produced up to 21 percent invented references. Forcing every citation through a lookup produced none.
• LightOn measured 87.69 against 64.98 on a long document test, holding the model and the corpus fixed and changing only the retrieval design.
• A bigger context window does not replace a lookup. Anthropic describes attention as a finite budget, with degradation that is gradual rather than a cliff.
• The embedding ranking has a separate Portuguese board, and the leader of the multilingual board does not appear on it.

What a market research knowledge base actually holds

PERSONAL EXPERIENCE A market research knowledge base is an indexed store of everything a research operation has already produced. That means transcripts, coded verbatims, tabulation runs, topline decks, reports and the briefs behind them. Indexed means a question can reach the passage, not only the file.

That last word does the work. Most archives are stores without an index. They hold twenty years of material, searchable by filename, client and date, which are the three things nobody asks about. The questions people bring have a different shape. Has this segment said this before. Which of our studies contradicts the one on the table now. Those questions cross studies, and they land inside documents rather than on their covers.

How a research team finds an old study today

PERSONAL EXPERIENCE Answer first: by asking someone who was there.

1. Someone remembers roughly when the study ran and who led it.
2. That person is asked by message and searches their own drive.
3. If they have left, the search moves to a shared folder organized by client and year.
4. Whoever is looking opens decks one by one until something matches.
5. The find is pasted into the new deck without the page reference.

Step five compounds the damage. The finding survives, the source does not, and next year the search starts again.

What the unfindable archive costs

The cost lands twice, and only the second one gets counted.

The first is the hours spent looking, which everyone sees and nobody logs. The second is larger and invisible: the study commissioned because nobody could prove the answer already existed. A client pays for fieldwork twice, years apart, on a question already answered.

PERSONAL EXPERIENCE The sentence we hear is the client's own: we still keep everything in folders, and we still ask the person who ran it. The archive is treated as storage rather than as an asset, and it belongs to the knowledge manager when that role exists at all.

Why a bigger context window does not solve it

Answer first: because attention is a budget, and filling it spends the budget.

A context window is the amount of text a model can consider at once. The temptation with twenty years of archive is to paste more of it in. Anthropic's engineering write up describes that window as a finite attention budget, degrading gradually rather than falling off a cliff. There is no threshold number to design against.

There is also a position effect. When the critical fact sits mid window rather than at either end, measured performance drops. The long context literature reports falls of 20 to 50 percent depending on the task, a range wide enough to read as a warning rather than a constant.

So the question was never how much fits. It is what deserves to be in there for this question, which is what a lookup decides.

The bottleneck moved from knowing to retrieving

This is the finding that reframes the whole problem.

Google Research reports that frontier models, measured as a group rather than one by one, hold 95 to 98 percent of the facts they were tested on. They fail to retrieve 26 to 34 percent of those when answering. The knowledge was acquired. It was not reached.

Read that across to a research archive. An operation with an index that still cannot get answers out of it has a retrieval problem, and retrieval is a design decision rather than a model choice. Switching models does not move a number that was never about the model.

Why every citation has to come from a lookup

Answer first: because a citation the model composes is a citation it can invent.

In the same Google Research work, a system that let citations come out of the model's memory produced up to 21 percent invented references. A system that required every citation to come from a retrieval produced none. That number describes the arrangement the source measured, not any product on the market, and not ours.

Provenance means each claim in an answer points back to the exact passage it came from, in a named document, on a findable page. It is the discipline we argued for in [keeping the evidence chain in qualitative coding](/blog/qualitative-coding-evidence-chain), moved up to the archive.

The working version is a rule an engineer enforces. A sentence with no retrieved passage behind it does not ship. That is architecture rather than prompting.

What the retrieval arrangement actually looks like

How a question reaches a passage in the archive

A left to right flow in five stages. Stage one, the archive: twenty years of transcripts, decks, tabulation runs and reports. Stage two, indexing: every document is split into passages, and each passage is written into two indexes at once, a word index that matches the exact terms a researcher typed and a meaning index that matches passages phrased differently. Stage three, the question: the same question is sent to both indexes and each returns its own candidate passages. Stage four, reordering: a second pass reads the question against each candidate and reorders the merged list, which is where the hard sets gain 5 to 15 points on the position of the first correct answer. Stage five, the answer: only the top passages are handed to the model, and every sentence in the answer points back to the passage and the document it came from. A note at the bottom states the rule that a sentence with no retrieved passage behind it does not ship.

Two searches feed one list, the list is reordered, and only then does anything reach the model. The citation is a byproduct of that route rather than something the model composes.

The arrangement that became standard in 2026 combines word matching with meaning matching and closes with a reordering pass. An embedding is the numerical form of a passage that lets a search find text meaning the same thing in other words. A store built only on embeddings misses the exact term a researcher typed, such as a brand name or a coded label.

Reordering, also called reranking, is a second pass that reads the question against each candidate passage and re-sorts the list before anything reaches the model. The retrieval literature puts the gain at 5 to 15 points on the metric tracking where the first correct answer lands, concentrated on hard collections. On easy ones it changes almost nothing, so it is a choice rather than a default. The reference board is MTEB.

Long documents change which design wins

UNIQUE INSIGHT A research archive is a long document archive, and that changes which design wins.

Four measured numbers, and who measured each one

Two bar panels. The top panel is titled what the model holds against what it hands back. A first horizontal bar shows 95 to 98 percent of tested facts stored inside frontier models. A second, much shorter bar below it shows 26 to 34 percent of those same facts not retrieved at the moment of answering. The owner line names Google Research, measuring frontier models as a group rather than one model. The bottom panel is titled long document retrieval with the model and the corpus held fixed. A first bar shows 64.98 for the single vector design. A second, longer bar shows 87.69 for the late interaction design. The owner line names LightOn, in a controlled test setup where only the retrieval paradigm changed, in the training languages of the model. Below both panels, two footnotes: adding a reordering pass gains 5 to 15 points on the position of the first correct answer, on hard collections only; and forcing every citation through a lookup took invented references from up to 21 percent down to none, in the specific arrangement the source measured. A closing line states that every number belongs to the source named beside it and none of them is a Cassi.ai measurement.

None of these are our measurements. Each one carries the owner it came from, and each one stops where its own base stops.

LightOn ran a controlled test setup on long documents, holding the base model and the corpus fixed and varying only the retrieval design. The late interaction approach compares the question against many points inside a passage rather than one summary number for the whole passage. It scored 87.69 against 64.98 for the single vector approach, in the languages the model was trained on.

That result should change a build decision, and it should stop short of two claims it cannot make. It says nothing specific about Portuguese, because the Portuguese note in that work sits inside an image and was never confirmed in the text. And one controlled comparison is not a promise about every archive.

A third design is worth knowing. A published survey of graph based retrieval maps the case where the answer depends on connecting information held in different documents, which is the cross study question. The survey maps the territory rather than declaring graph better than vectors, and neither do we.

Reading the embedding rankings for a Portuguese archive

Answer first: the general ranking is the wrong ranking for a Brazilian client.

The MTEB leaderboard publishes a separate board for Portuguese, and the model leading the multilingual board does not appear on it. For an archive of Brazilian transcripts and decks, picking by the headline board is picking against your own corpus.

Two more rules come from the ranking's own notices. The newer benchmark version is not comparable with the previous one, so a number from one quoted beside a number from the other is invalid. And the project removed its private test column in January 2026, after the company that co-developed the test turned out to hold the evaluation data.

Where Cassi.ai comes in

Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights, working with research agencies and corporate insights teams. The product here is the Insight Lab, and it closes a gap clients name themselves: there is no library with the house's own files indexed, so the assistant everyone uses starts from zero at every question.

What we build is the route in the first figure, sized to the client's archive. Passages instead of files, two searches instead of one, a reordering pass, and an answer where every claim points at its document and page. It runs wherever the contract says it must, the argument in [open weight models and what the license allows](/blog/open-weight-models-in-market-research). The rest is in the Cassi.ai portfolio and in [the six gaps a chat assistant leaves open](/blog/why-chatgpt-is-not-a-research-platform).

When indexing the archive is not worth it

Retrieval numbers behind this table are Google Research's and LightOn's. The mapping onto research operations is ours.

FAQ

What is a market research knowledge base?

An indexed store of everything a research operation has produced: transcripts, coded verbatims, tabulation runs, decks, reports and briefs, searchable at the passage rather than the file. Ask a question that crosses three studies and see whether the answer arrives with page references.

Why does a larger context window not replace retrieval?

Because attention is a budget rather than a container. Anthropic describes degradation as gradual rather than a cliff. The long context literature reports drops of 20 to 50 percent, by task, when the critical fact sits mid window. Deciding what goes in beats making the window bigger.

How do I stop an assistant from inventing sources?

Require every citation to come from a retrieval. Google Research reports up to 21 percent invented references when citations may come from the model's memory, and none when each comes from a lookup. That measurement describes the arrangement the source built. Treat it as a design principle rather than a number to quote for your own system.

Which embedding model should I use for a Portuguese archive?

Not the one at the top of the multilingual board. The MTEB leaderboard keeps a separate Portuguese board and the multilingual leader does not appear on it. Test on your own documents, with your own questions, and record which benchmark version you are reading. Versions are not comparable with each other.

Is a graph based design better than vector search?

Unproven either way. The published survey maps graph retrieval as an option when the answer spans several documents. It maps rather than ranks.

An archive nobody can search is a storage cost with a good story attached. What decides whether yours is more than that is how often last year somebody commissioned work the archive had already answered.

If that count is above zero, take three questions from last quarter, answer them from the archive alone, and time each one. Whatever takes longer than ten minutes is what the index is for.

Published by Cassi.ai. Read the full article at https://www.cassiai.com/blog/market-research-knowledge-base.