Contextual Retrieval
A RAG optimization approach proposed by Anthropic in 2024. It uses a large language model to generate an independent contextual summary for each document chunk, which is prepended before encoding. This approach addresses the degradation in retrieval quality caused by chunks being processed without surrounding context.
Core Facts
Timeline (last 90 days)
In its official article on Contextual Retrieval, Anthropic gives the example that an embedding model may find generic content about an error code yet miss an exact "TS-999" match.
Contextual Retrieval means that before indexing each chunk, a large model is first used to generate a context prefix explaining that chunk's position in the whole document, which is then indexed together with the chunk for embedding and BM25.
Anthropic's experiments showed that combining embedding with BM25 and adding context to chunks reduced the top-20 retrieval failure rate by 49% (from 5.7% to 2.9%).
The Contextual Retrieval approach proposed by Anthropic in its technical blog can reduce retrieval failure rates by approximately 49%.
Anthropic proposed the Contextual Retrieval approach in 2024, which uses a large language model to generate a contextual summary for each chunk before concatenating and encoding it.
21.2% is a relative reduction rather than an absolute one, meaning the new version eliminates approximately one-fifth of the failures relative to the earlier version's failure rate baseline.
Contextual Retrieval was proposed by Anthropic; it prepends an LLM-generated contextual summary to each chunk so that the chunk retains the necessary background information even when separated from the original text.
All Facts (7)
Contextual Retrieval was proposed by Anthropic; it prepends an LLM-generated contextual summary to each chunk so that the chunk retains the necessary background information even when separated from the original text.
65%UnverifiedIn its official article on Contextual Retrieval, Anthropic gives the example that an embedding model may find generic content about an error code yet miss an exact "TS-999" match.
50%UnverifiedContextual Retrieval means that before indexing each chunk, a large model is first used to generate a context prefix explaining that chunk's position in the whole document, which is then indexed together with the chunk for embedding and BM25.
50%UnverifiedAnthropic's experiments showed that combining embedding with BM25 and adding context to chunks reduced the top-20 retrieval failure rate by 49% (from 5.7% to 2.9%).
50%UnverifiedThe Contextual Retrieval approach proposed by Anthropic in its technical blog can reduce retrieval failure rates by approximately 49%.
50%UnverifiedAnthropic proposed the Contextual Retrieval approach in 2024, which uses a large language model to generate a contextual summary for each chunk before concatenating and encoding it.
50%Unverified21.2% is a relative reduction rather than an absolute one, meaning the new version eliminates approximately one-fifth of the failures relative to the earlier version's failure rate baseline.
50%