nvidia/nemotron-3.5-embed-1b

{"about":"### Semantic search across text\n\nNemotron 3.5 Embed 1B converts text into a numerical representation of its content. This NVIDIA model is intended for document search by meaning: a question and a relevant passage may use different wording. Such representations support search across reference material, internal instructions, articles and support answer collections.\n\nThe underlying NVIDIA model combines text and document images in a shared representation space; generation 3.5 adds retrieval based on the visual content of pages. This Ailoot API listing uses its text workflow: text input and vector output.\n\nIn Ailoot API, the model accepts text through the embeddings method and returns a vector. The vector is used to compare a query with documents; the next application stage composes the final answer for the user. This lets you organize source retrieval separately from answer preparation based on those sources.\n\n### How to build a search workflow\n\nFirst divide the material into self-contained, meaningful chunks and obtain an embedding for each. Store each vector alongside its source text, document title and identifier. At search time, convert the question using the same model and retrieve similar vectors from the index. The resulting passages can be shown to the user or passed to a text model as context for a document-grounded answer.\n\nQueries and documents use separate representations, which are then compared within a single search workflow. When preparing the corpus, keep a statement together with its explanatory context so the retrieved chunk can be read independently and connected to its source document.\n\n### Model characteristics\n\nThe 1B version was developed using the larger Nemotron 3.5 Embed 8B as a teacher model. Its text component is based on Nemotron 3 Embed 1B and retains that model’s text representation space. The underlying model’s description specifies a vector of 2,048 values with L2 normalization. Consider this when designing the index, checking the dimensionality of the selected method’s actual response.\n\nText retrieval was evaluated in 34 languages, including Russian. This coverage makes the model a candidate for a multilingual knowledge base; evaluate retrieval quality using real questions and documents from your subject area.\n\n### Input and request size\n\nThe current Ailoot API text method sets a limit of 4,096 tokens per input. Index longer material as separate chunks with their own source references. This listing’s available workflow is text input and embedding output.","alias_of":null,"aliases":[],"author":"NVIDIA","canonical_model":"nvidia/nemotron-3.5-embed-1b","ctx":4096,"formats":["embeddings"],"id":"nvidia/nemotron-3.5-embed-1b","input":["text"],"kind":"embedding","max_out":0,"methods":[{"cache_read":null,"cache_write":null,"cache_write_1h":null,"cache_write_5m":null,"conditional_meter":null,"ctx":4096,"family":"openai","fast_multiplier":null,"flex_multiplier":null,"fmt":"embeddings","http":"POST","img_in":null,"img_out":null,"in":0.1,"input":["text"],"label":"Эмбеддинги","max_out":0,"max_tool_calls":null,"out":0,"output":["embedding"],"path":"/v1/embeddings","request_model_id":"nvidia/nemotron-3.5-embed-1b","route_label":"Core","status_reads":[],"tiers":[],"tool_prices":null,"unit":"token","units":[]}],"model_aliases":[{"canonical_model_id":"nvidia/nemotron-3.5-embed-1b","request_model_id":"nvidia/nemotron-3.5-embed-1b","route_label":"Core"}],"name":"nvidia/nemotron-3.5-embed-1b","price":{"in":0.1,"in_label":null,"out":null,"out_label":null,"sort_in":0.1,"sort_out":null,"varies":false},"public_id":"nvidia/nemotron-3.5-embed-1b","short":"NVIDIA text embeddings for multilingual semantic search and source retrieval in RAG."}