Grounding a model in your own data: embeddings, search, chunking, RAG.
Sign in to track your progress across sections.
Turning text into vectors, and the store that finds the nearest ones.
6 lessons · 20 min · Retrieval and knowledge
Chunking, storage, retrieval and generation, and the failure each stage introduces.
8 lessons · 35 min · Retrieval and knowledge
Split the same document by size, by structure and by section, then compare what retrieval returns.
90 min · Retrieval and knowledge
Upload a set of documents, ask questions about them, and get answers that cite what they came from.
8–12 hours · Retrieval and knowledge
44 articles
Listed inEmbedding Models on Hugging FaceRetrieval and knowledgeon
Benchmark scores across retrieval, clustering, and reranking tasks. Useful for a shortlist, dangerous as a final answer.
Listed inSentence TransformersRetrieval and knowledgeon
The library most self-hosted embedding pipelines are built on: pooling, training, and inference.
Listed inOpenAI Embeddings APIRetrieval and knowledgeon
Shortening a vector is supported and cheap, and slicing it yourself without normalizing degrades search with nothing failing.
Listed inAnomaly DetectionRetrieval and knowledgeon
An anomaly score is a distance read backwards. Plus the unnormalized-vector bug that makes a detector flag your longest documents.
Listed inEmbedding Models on Hugging FaceRetrieval and knowledgeon
Changing embedding models means re-embedding the whole corpus, so the choice is harder to reverse than a generation model. What to check on a model card: vector dimensions, maximum sequence length, licensing, and whether the model expects a prompt prefix that fails silently when omitted.
Listed inRecommendation SystemsRetrieval and knowledgeon
Nearest neighbours of an item are a recommender you already have. It measures whether two things are about the same subject, which is not the same as predicting what a person will engage with.
Listed inJina EmbeddingsRetrieval and knowledgeon
A 32,768-token context means the encoder stops dictating your chunk size. Task types move the query-versus-passage distinction into a parameter a reviewer can see, and truncatable vectors have one way to cut them safely.
Listed inWhat are Embeddings?Retrieval and knowledgeon
A vector of floating point numbers whose distance from another measures how related two texts are. The same distance reads as search, recommendation, classification or anomaly depending on the question you ask it.
Listed inCohere EmbedRetrieval and knowledgeon
`input_type` is the parameter you cannot skip: your corpus takes `search_document` and your queries take `search_query`, and getting it backwards ranks worse with no error to tell you. Binary compression takes a 1024-dimension vector to 128 bytes.
Listed inEmbedding ModelsRetrieval and knowledgeon
Vectors from two models are not comparable, so the choice comes with a migration attached from the day you make it. Dimensions, input limits, and the mixed index that fails without erroring.
Listed inData ClassificationRetrieval and knowledgeon
Embed once and compare each item to a class centroid, which costs one cheap call where prompting costs an expensive one forever. The distance is also a confidence signal, so you can route the ambiguous cases and say when something fits no class at all.
Listed inSentence TransformersRetrieval and knowledgeon
Embeddings on your own hardware, which settles the question for data that cannot leave your network. `CrossEncoder` is the class most RAG systems are missing: retrieve widely with a bi-encoder, then rerank the top results with a model that reads query and document together.
Listed inSemantic SearchRetrieval and knowledgeon
Retrieving by meaning finds the document that shares no words with the query, and loses to keyword search on identifiers, rare terms and quoted phrases. Hybrid search runs both and fuses the results.
Listed inGemini EmbeddingRetrieval and knowledgeon
Embedding 2 dropped the `task_type` parameter, so task specification moved into the prompt where a typo is a silent quality change rather than an error. It also returns one aggregated vector for multiple inputs unless each is wrapped separately.
Listed inPerforming Similarity SearchRetrieval and knowledgeon
Cosine and dot product rank the same on normalized vectors. The decision that actually changes your results is where the filter runs.
Listed inSupabase VectorRetrieval and knowledgeon
Vectors as a column, so the filter is a WHERE clause on columns you already have and a row and its embedding land in one transaction.
Listed inIndexing EmbeddingsRetrieval and knowledgeon
Qdrant's explainer on HNSW, payload indexes, and the quantization trade-offs behind recall and memory.
Listed inPurpose and FunctionalityRetrieval and knowledgeon
A B-tree cannot find nearest neighbors, exact search stops scaling, and approximate search trades a guarantee for speed. Sometimes you need none of it.
Listed inLanceDBRetrieval and knowledgeon
An embedded retrieval library rather than a server, with multimodal assets beside the vectors. Plus what the project does not actually claim.
Listed inWeaviateRetrieval and knowledgeon
A search service rather than a vector column or an embedded library. Hybrid search fuses keyword and vector ranking, which recovers the exact identifiers embeddings lose, and the index types include a flat one suited to many small per-tenant datasets.
Listed inChromaRetrieval and knowledgeon
The persistent client gets a working query with no infrastructure decision, and the collection code survives the move to a server. Concurrency, the directory's lifetime and backups are the bill that arrives with traffic.
Listed inMongoDB Atlas Vector SearchRetrieval and knowledgeon
`$vectorSearch` runs as the first stage of an aggregation pipeline, so retrieval composes with the joins and grouping that follow it. Exposing both ANN and exhaustive search in one system means you can measure what your recall settings actually return.
Listed inIndexing EmbeddingsRetrieval and knowledgeon
Recall, latency and memory are the three quantities every index decision trades between, and knowing which one you are short of is most of the work. Each family gives you one build-time parameter and one runtime recall dial, `nprobe` for IVF and `efSearch` for HNSW.
Listed inQdrantRetrieval and knowledgeon
Filters run during the search rather than after it, which is a correctness feature dressed as an implementation detail. Post-filtering silently returns fewer results than you asked for, and in a multi-tenant system it passes every test until it does not.
Listed inVector DatabasesRetrieval and knowledgeon
Purpose-built stores for nearest-neighbour search over millions of vectors.
Listed inFAISSRetrieval and knowledgeon
A similarity-search library with no server, no filtering and no persistence you did not write, sitting a layer below the vector databases that use it. Right when you already have the storage and want control over the index, and wrong when you would end up rebuilding a database badly.
Listed inPineconeRetrieval and knowledgeon
The store where nobody on your team runs anything, and namespaces are the feature to build around. A per-tenant namespace is a hard partition rather than a filter you remembered to apply, which removes a class of leak by construction.
Listed inChunkingRetrieval and knowledgeon
Structural, recursive and semantic splitting compared, with the sizes to start from for short lookups and for prose. Chunk size sets a ceiling on what retrieval can return, and several common mistakes lower it further.
Listed inChunkingRetrieval and knowledgeon
Prepending model-generated context to each chunk before embedding, and the measured drop in failed retrievals it produced.
Listed inChunkingRetrieval and knowledgeon
Pinecone walks through fixed, recursive, and document-aware splitting with concrete size and overlap recommendations.
Listed inLangChainRetrieval and knowledgeon
LangChain's concept doc on the splitter family — what each one optimises for and when to reach for it.
Listed inStoring in a Vector DBRetrieval and knowledgeon
A namespace is a hard partition, not a tag. Changing the embedding model invalidates every vector, so plan the rebuild before you need it.
Listed inRAGFlowRetrieval and knowledgeon
Parse the layout before chunking, because a flat extractor turns a pricing table into numbers with no headers and nothing downstream can repair it.
Listed inChunkingRetrieval and knowledgeon
Splitting documents so retrieved passages are both findable and self-contained.
Listed inLlamaIndexRetrieval and knowledgeon
Documents and Nodes, and a query stage that separates retrieval, postprocessing and response synthesis. Postprocessing is the step hand-written pipelines usually skip, and it is where reranking and filtering belong.
Listed inRAG Use CasesRetrieval and knowledgeon
Four problem shapes retrieval fits and four it does not, why counting questions fail quietly, and how a million-token window moved the threshold.
Listed inRAG with Dynamic FiltersRetrieval and knowledgeon
A forgotten metadata filter returns everyone's documents; a forgotten namespace returns an error. When the filter is a tenancy boundary, that asymmetry is the whole argument for a partition.
Listed inRetrieval ProcessRetrieval and knowledgeon
Anthropic's evaluation retrieves 150 candidates and reranks to 20. Why the two stages are good at different things, how to choose each number, and the measurement that separates retrieval from generation.
Listed inEmbedding DocumentsRetrieval and knowledgeon
Prepending context to a chunk before embedding it cut Anthropic's top-20 retrieval failure rate from 5.7% to 3.7%. What to store beside the vector, and why ingest has to be safe to run twice.
Listed inHaystackRetrieval and knowledgeon
Pipelines are a declared graph, so branching a query into keyword and vector retrieval and merging the results is something you can draw and inspect. The structure pays at five stages and gets in the way at two.
Listed inRAG vs Fine-tuningRetrieval and knowledgeon
Retrieval changes what a model can see and fine-tuning changes how it behaves, which settles less than it sounds. Four questions decide the case, and the production answer is usually both, aimed at different jobs.
Listed inLangChainRetrieval and knowledgeon
Two things arrive together: a common interface over providers, and a framework for composing calls. The recurring complaint is not that it fails but that it obscures, since finding what actually reached the model takes longer through code that assembled it.
Listed inWhat is RAG?Retrieval and knowledgeon
Two pipelines rather than one: indexing when documents arrive, querying per request. Each of the four stages introduces its own failure, and retrieval always returns something even when your corpus cannot answer.
Listed inGenerationRetrieval and knowledgeon
The last stage is where retrieval gets believed. Separating passages from instructions, citations that carry a validity guarantee rather than a plausible ID, and building refusal for the question your corpus cannot answer.