In CrewAI, how do you attach a PDF or text file as crew knowledge?

## What "knowledge" means in CrewAI Knowledge is the grounding material you attach to a crew *before* it runs: documents, records or raw strings that the agents should be able to consult while doing their work. It is distinct from tools (which fetch data at call time) and from memory (which accumulates from the run itself). You declare it once, CrewAI indexes it, and agents retrieve from it automatically as part of task execution. ## The source classes Each kind of input has its own class, imported from `crewai.knowledge.source.*`: - `StringKnowledgeSource(content="...")` — inline text, useful for policies or facts you generate in code. - `TextFileKnowledgeSource(file_paths=["notes.txt"])` - `PDFKnowledgeSource(file_paths=["report.pdf"])` - `CSVKnowledgeSource(file_paths=["sales.csv"])` - `JSONKnowledgeSource(file_paths=["config.json"])` - `ExcelKnowledgeSource(file_paths=["book.xlsx"])` - `CrewDoclingSource(file_paths=[...])` — routes the file through Docling, which handles a broader set of document formats and preserves more structure than the plain readers. All of them accept `chunk_size` and `chunk_overlap`, so you can tune how the document is cut up without writing your own splitter. ## The `knowledge/` directory rule File-based sources do **not** take arbitrary filesystem paths. CrewAI resolves `file_paths` against a `knowledge/` directory at the root of your project. So `PDFKnowledgeSource(file_paths=["handbook.pdf"])` looks for `<project-root>/knowledge/handbook.pdf`, and `file_paths=["docs/handbook.pdf"]` looks for `<project-root>/knowledge/docs/handbook.pdf`. Passing an absolute path to a file outside that tree, or leaving the PDF beside your `main.py`, produces a file-not-found error at kickoff. This trips up almost everyone once; remembering it is worth an easy interview point. ## Crew level versus agent level `Crew(knowledge_sources=[...])` makes the material available across the crew. `Agent(knowledge_sources=[...])` scopes it to that one agent, which matters when different roles should not see each other's material (a legal reviewer's contracts versus a marketer's brand guide) and when you want to keep each agent's retrieval set small and on-topic. You can use both together. ## What happens at kickoff When the crew starts, CrewAI takes each source, produces its chunks, embeds them with the configured embedder (OpenAI embeddings by default, which is why an `OPENAI_API_KEY` is required unless you override `embedder`), and writes them into the knowledge store on disk. During task execution the agent's query is embedded and the closest chunks are pulled back and injected into the prompt as extra context. Because the indexing happens on the way in, the very first kickoff after adding a large document is slower and costs embedding tokens; subsequent runs reuse what is already stored. ## Tuning retrieval If agents get too much or too little grounding, the knobs are the source's `chunk_size`/`chunk_overlap` and a `knowledge_config` on the agent or crew, built from `KnowledgeConfig`, which controls how many results come back and the minimum similarity score they must clear. Smaller chunks give sharper matches but lose surrounding context; larger chunks carry context but dilute the embedding and eat prompt budget. ## Common mistakes Attaching a 300-page PDF and expecting the agent to "read" it — it never sees the whole thing, only the retrieved chunks. Assuming knowledge updates itself when the file changes — you generally need to reset the knowledge store so the new content is re-embedded. And confusing knowledge with tools: if the data must be current at call time (a live API, a database query), that is a tool, not a knowledge source. ## Interview framing A strong answer names two or three concrete source classes, states the `knowledge/`-relative path rule, mentions both the crew-level and agent-level attachment points, and finishes by saying what happens mechanically — chunk, embed, store, retrieve into the prompt — rather than describing knowledge as something the model "has read".