FINDING DIRECTIONContext can give words their meaning, but when part of a conversation is missing, the listener will fill the empty space with an explanation of their own.
In the Three's Company episode “Chrissy's Hospitality”, Mr. Furley stands in the living room while Jack and Chrissy work behind the closed bathroom door. He can hear them, but he cannot see what they are doing.
Jack: Okay, Chrissy, I'll get in the tub with you, then we can get it on.
Chrissy: Get next to me. I'll show you what to do.
Jack: This isn't exactly the first time I've ever done this.
Chrissy: Maybe so, but girls are better at this than boys.
Jack: Come on, Chrissy. A little less talk and a little more action.
Chrissy: Okay, you do your part, I'll do mine. I don't think it'll reach.
Jack: Of course not. You've got to unfold it first.
Janet comes out of the kitchen and sees Mr. Furley gawking in the living room.
Janet: Mr. Furley, what are you doing?
Mr. Furley: Never mind what I'm doing. What are they doing?
Janet: Nothing.
Mr. Furley: Oh, don't give me that. I heard! Rub-a-dub-dub, come hop in my tub. I heard all of it!
Janet: You heard all of what?
Jack comes out of the bathroom.
Jack: What is going on in here?
Mr. Furley: How'd you get your clothes on so fast?
Jack: What?
Mr. Furley: Don't act so innocent. I know what's what! How'd you get your clothes on so fast?
Jack: What is he talking about?
Janet: I don't know. Rub-a-dub-dub, come hop in my tub?
Jack: What are you talking about?
Chrissy comes out of the bathroom.
Chrissy: Jack, are you going to help me or not?
Mr. Furley: Aha! How'd you get your clothes on so fast? Another innocent. I know what you were doing.
Chrissy: Putting up a shower curtain.
Mr. Furley: That's right. You were—putting up a shower curtain?
When you supply a model with a 200-page manual and ask a question about one paragraph on page 143, the same misunderstanding can happen. The model may have the entire file, retrieve a few passages from an index, or use a search tool to find the relevant section. Each approach gives the model different context for the same question.
[ FIELD NOTE / ATTACHMENTS ]
Attaching a file to a conversation does not guarantee the model received the context you are asking about.
The Working Material
Anthropic defines context as the set of tokens included when a language model generates a response. Those tokens can come from several places, and each has a different job.
- Instructions define the model's role, constraints, priorities, and expected output.
- Conversation carries the current request and the exchanges that led to it.
- Sources provide documents, code, images, records, and other evidence needed for the task.
- State records what an agent has already tried, changed, decided, or learned from a tool.
The context window is the model's working material for that moment. A file stored elsewhere may be available to the application without being present in that window. The application still has to select it, retrieve it, or let the agent find it.
This makes context a product decision. The model's answer depends partly on what the interface and the agent harness placed in front of it.
What Happened to My File?
Small files can sometimes be included in full. That gives the model every line, but the file has to share a finite context window with instructions, conversation history, tool definitions, and other sources.
Larger collections usually require retrieval. A common process looks like this:
The application extracts the file's text and divides it into smaller chunks. It can create an embedding for each chunk, which represents the passage in a form that supports semantic similarity. It may also build a keyword index for exact terms. When a question arrives, the retrieval service searches those indexes, ranks the results, and adds a limited number of passages to the model's request.
Anthropic's description of retrieval-augmented generation shows how semantic and keyword searches can be combined before the highest-ranking chunks are added to a prompt. OpenAI's vector-store search similarly returns relevant chunks with filenames, attributes, and similarity scores. Its file ingestion API also exposes the strategy used to divide a file into chunks.
Indexing turns the document into pieces designed to be found later, separating each piece from some of the continuity of the original.
[ FIELD NOTE / RETRIEVAL ]
Access to a file and attention to its contents are separate conditions.
What Got Lost
A chunk can contain the right sentence without enough information to identify what the sentence describes. “Coverage begins after 90 days” is useless when the product, customer, jurisdiction, and effective date were established several paragraphs earlier.
Chunk overlap can preserve some nearby text. Headings, filenames, and metadata can restore part of the missing identity. Better retrieval can combine exact words with semantic matches and rerank the candidates before sending them to the model. None of these steps guarantees that the required passage will be selected.
Indexing also introduces time. A file can change after its chunks were created. Two versions can remain in the same collection. A scanned table may parse badly. A code index may preserve the function body while losing the types or configuration that determine its behavior.
Agents have another option. Instead of relying entirely on a prepared index, they can keep file paths, links, and search tools available, then inspect sources as the work requires them. Anthropic describes this as just-in-time context: the agent uses tools such as file search and grep to load relevant material at runtime. This can reach current files and follow relationships that an earlier indexing pass did not preserve.
File names, repository structure, instructions, and tool descriptions guide where the agent looks.
What Matters
Every added token occupies part of a finite context window. Anthropic notes that models can become less precise at retrieving information and reasoning across distant passages as the context grows. Useful context has a clear relationship to the current decision.
The source should identify itself and remain easy to check. Keep the document title, section, date, version, and location with each retrieved passage. Structured headings and descriptive filenames give both retrieval and the model more information to work with. Citations, passages, file paths, and tool history show if the model used the right information.
Mr. Furley received every word he needed to draw his own conclusion. What he did not see was the thing that made the conversation ordinary. Retrieval can create the same problem by returning the words and leaving their subject behind.
When AI confidently explains what happened in the room, check whether its context included the shower curtain.