Full-picture answers
Full-picture answers from your company knowledge
Ask a question and get the whole answer. Most AI tools keep showing you the same document, so whatever your other documents say never makes it into the answer. Tricky Wombat includes every different point your company has on the question and links each point to the document it came from, so nothing gets missed.
The definition
A full-picture answer covers every side of the question
Ask a question with more than one side to it and your documents will hold several distinct ideas about it. A policy document says one thing. An architecture note adds a constraint. A Slack thread records the exception someone agreed to last quarter. A full-picture answer includes all of them, with a citation back to where each came from.
A partial answer is what you get when the system reads its closest matches and stops. It repeats the idea that appears most often in your files and never reaches the ones that live in a single document. Tricky Wombat reads what each passage is about before it writes, so every idea is present in the reply.
- Every distinct idea in your documents shows up in the answer
- The same point is never repeated, and each one links to the file it came from
- An idea found in one document gets the same place as an idea found in twenty
The failure
Why AI search repeats one document and misses the rest
Standard AI search scores every passage in your knowledge base against your question and keeps the ten highest. That rule has a blind spot. If seven of the ten best-scoring passages are the same compliance paragraph copied into seven documents, the model reads that paragraph seven times and the other three ideas never make it in.
Order and clutter also change the answer. Stanford researchers found that model accuracy drops by more than 30% depending on where a fact sits in what the model reads, even when the fact is present. Chroma tested 18 frontier models and found that near-duplicate passages were the content most often tied to made-up answers. Enterprise search tools get the right answer on the first try about 10% of the time, against roughly 95% for a consumer Google search.
- Scoring passages one at a time rewards repetition, so copies crowd out variety
- Facts buried in the middle of a long, repetitive context get missed
- Near-duplicate passages are the content most often linked to invented answers
How it works
Pull wide, sort by idea, keep one of each
Tricky Wombat starts by pulling more than it needs, around fifty passages instead of ten. It then compares those passages to each other rather than to the question, and groups the ones that say the same thing. Two files that use different words for the same idea land in the same group. One file that covers two ideas contributes to two groups. It is what you would do with a stack of printouts: sort them into piles by what they are about, then read the top sheet from each pile.
From each group, one passage goes forward to the model. The groups are built the same way every time, so the same question returns the same documents every time. That is what makes the answer repeatable and the documents easy to check. The sorting step takes a few milliseconds, so you never wait for it.
- Fetch about fifty candidate passages instead of ten
- Group passages that mean the same thing, even when the wording differs
- Send the model one passage per idea instead of seven copies of the same one
- Same question, same documents, same groups, every time
The payoff
Less for the model to read, fewer invented answers, nothing dropped
Published tests of this approach cut the text sent to the model by 46% to 90% with no loss in answer quality across GPT-4, Llama 2, and Mixtral. A separate study on conversational search found that grouping related passages and placing them together improved answer quality by up to 54% over unordered context. Less text costs less to run and gives the model fewer places to go wrong.
The business case is the shape of real questions. "What is our approach to data residency for EU customers?" needs a legal policy, an architecture diagram, a customer FAQ, and an internal thread about a regulatory change. Ten closest matches tend to return four versions of the legal policy and none of the architecture. One passage per idea returns all four.
- 46% to 90% less text sent to the model, same answer quality in published tests
- Up to 54% better answer quality when related passages are grouped and placed together
- Multi-part questions get all their parts answered in one reply
Balance
The most-copied document stops drowning out the one that matters
In most companies the most useful knowledge is the least repeated. The standard policy is pasted into fifty documents. The engineer's note explaining why the policy has an exception exists in one file. Search that ranks by similarity and keeps the top ten returns the policy ten times and never shows the note. The Economist called this the tyranny of the majority in AI research tools: the common answer wins because it is common.
Grouping by idea reverses that. Fifty copies of the policy collapse into one group and one passage. The engineer's note forms its own group and gets its own passage. You also see the groups that fed the answer, so when a reply surprises you, you can check which ideas it drew on and which file each came from.
- Fifty copies of the same paragraph count as one idea, not fifty votes
- A specialist note found in only one file still reaches the model
- You can see which ideas and which files were behind each answer
Side by side
Full-picture answer versus a top-matches answer
A top-matches answer is what most AI search produces: the ten closest passages, read in order. A full-picture answer is what Tricky Wombat produces after sorting those passages by idea.
Comparison as of October 1, 2026
| Capability | Full-picture answer | Top-matches answer |
|---|---|---|
| Every distinct idea in your documents is represented | Yes | Partial1 |
| Duplicate passages removed before the model reads them | Yes | - |
| A point made in one document gets the same weight as one copied into fifty | Yes | - |
| The same question returns the same documents every time | Yes | Partial2 |
| Less text sent to the model for the same question | Yes | - |
| You can see which ideas fed the answer as well as which files | Yes | - |
| Finds the passages closest to the question | Yes | Yes |
- 1.A top-matches answer covers every idea only when the ten highest-scoring passages happen to differ from one another. On repetitive company content, they usually do not.
- 2.Ranking is stable, but any change in which passage edges into tenth place changes what the model reads. Grouping by idea absorbs that noise because the group still sends one representative.
Legend: Yes = how Tricky Wombat answers · Partial = sometimes, depending on what ranked highest · - = not part of the approach. Top-matches answer = the standard retrieve-and-read pipeline. It is not a column for a named vendor.
Full-picture answers FAQ
The questions people ask about incomplete AI answers
An answer that includes every different point your documents make on a question, with a citation for each. Tricky Wombat produces one by pulling a wide set of passages, grouping the ones that say the same thing, and sending the model one passage per group. The term is ours. The method is agglomerative clustering applied between search and answer, which our technical page covers in depth.
Connect the folders and tools that hold the answer, then ask the question the way you would ask a colleague. Tricky Wombat pulls a wide set of passages from those documents, groups the ones that say the same thing, and writes from one passage per group. You get every idea, each with the file it came from. The newest policy folder on its own will not do it, because the exception and the diagram usually live somewhere else.
Most systems keep the ten passages that score highest against your question and read them in order. On company content, where the same paragraph is copied into many files, those ten are often several copies of one idea. The other ideas never reach the model, so the answer is accurate about one thing and silent about the rest.
A company question usually has several parts, and the closest passages often cover only one of them. Ask about data residency for EU customers and the highest-scoring passages may all be the legal policy, so the architecture constraint and the exception recorded in Slack never appear. Tricky Wombat groups the passages by idea and keeps one from each group, so each part of the question points back to a document in the reply.
Ask what happens after search returns its matches. If the system sends the ten closest passages straight to the model, copies of one paragraph fill the reply and the other ideas drop out. A complete answer needs a step that groups those passages by idea and keeps one of each before the model writes. That is the step Tricky Wombat runs.
Because that document, or copies of it, scored highest for your question and filled the available slots. Scoring passages one at a time rewards repetition. Grouping passages by idea before answering collapses the copies into one slot and frees the others for ideas found in only one document.
Search ranks passages by how closely they match the question, then keeps the top ten. A policy pasted into fifty files produces many high-scoring copies. The engineer's note that lives in one file produces one. The copies take the slots and the note never reaches the model. Grouping by idea collapses the fifty copies into a single slot, so the note keeps a place of its own.
No. Stanford researchers found that going from 20 documents to 50 improved accuracy by about 1.5% for GPT-3.5 Turbo, and that accuracy fell by more than 30% depending on where a fact sat in the context. More text does not help. Organized text does.
Better search, including the concept-level matching described on our Cognitive Resonance page, finds the right passages. Full-picture answers organize those passages after they are found: remove the duplicates, confirm each idea is present, and hand the model a set it can read well. Both steps matter. Most vendors invest only in the first.
No. Grouping fifty to a hundred passages takes single-digit milliseconds. The model then reads less text than it otherwise would, so the full reply usually arrives sooner.
See the full picture on your own documents.
Connect a folder and ask a question with more than one side to it.
Schedule a callTechnical deep dive: Agglomerative clustering · Cognitive Resonance: finding the right passages · What is a Company Brain