Chroma or ChromaDB is open-source data infrastructure tailored to applications with large language models. Its headquarters are in San Francisco. In April 2023, it raised 18 million US dollars as seed funding. ChromaDB has been used in academic studies on artificial intelligence, particularly as part of the tech stack for retrieval-augmented generation.
Scale space implementation
In the areas of computer vision, image analysis and signal processing, the notion of scale-space representation is used for processing measurement data at multiple scales, and specifically enhance or suppress image features over different ranges of scale (see the article on scale space). A special type of scale-space representation is provided by the Gaussian scale space, where the image data in N dimensions is subjected to smoothing by Gaussian convolution. Most of the theory for Gaussian scale space deals with continuous images, whereas one when implementing this theory will have to face the fact that most measurement data are discrete. Hence, the theoretical problem arises concerning how to discretize the continuous theory while either preserving or well approximating the desirable theoretical properties that lead to the choice of the Gaussian kernel (see the article on scale-space axioms). This article describes basic approaches for this that have been developed in the literature, see also for an in-depth treatment regarding the topic of approximating the Gaussian smoothing operation and the Gaussian derivative computations in scale-space theory, and for a complementary treatment regarding hybrid discretization methods. == Statement of the problem == The Gaussian scale-space representation of an N-dimensional continuous signal, f C ( x 1 , ⋯ , x N , t ) , {\displaystyle f_{C}\left(x_{1},\cdots ,x_{N},t\right),} is obtained by convolving fC with an N-dimensional Gaussian kernel: g N ( x 1 , ⋯ , x N , t ) . {\displaystyle g_{N}\left(x_{1},\cdots ,x_{N},t\right).} In other words: L ( x 1 , ⋯ , x N , t ) = ∫ u 1 = − ∞ ∞ ⋯ ∫ u N = − ∞ ∞ f C ( x 1 − u 1 , ⋯ , x N − u N , t ) ⋅ g N ( u 1 , ⋯ , u N , t ) d u 1 ⋯ d u N . {\displaystyle L\left(x_{1},\cdots ,x_{N},t\right)=\int _{u_{1}=-\infty }^{\infty }\cdots \int _{u_{N}=-\infty }^{\infty }f_{C}\left(x_{1}-u_{1},\cdots ,x_{N}-u_{N},t\right)\cdot g_{N}\left(u_{1},\cdots ,u_{N},t\right)\,du_{1}\cdots du_{N}.} However, for implementation, this definition is impractical, since it is continuous. When applying the scale space concept to a discrete signal fD, different approaches can be taken. This article is a brief summary of some of the most frequently used methods. == Separability == Using the separability property of the Gaussian kernel g N ( x 1 , … , x N , t ) = G ( x 1 , t ) ⋯ G ( x N , t ) {\displaystyle g_{N}\left(x_{1},\dots ,x_{N},t\right)=G\left(x_{1},t\right)\cdots G\left(x_{N},t\right)} the N-dimensional convolution operation can be decomposed into a set of separable smoothing steps with a one-dimensional Gaussian kernel G along each dimension L ( x 1 , ⋯ , x N , t ) = ∫ u 1 = − ∞ ∞ ⋯ ∫ u N = − ∞ ∞ f C ( x 1 − u 1 , ⋯ , x N − u N , t ) G ( u 1 , t ) d u 1 ⋯ G ( u N , t ) d u N , {\displaystyle L(x_{1},\cdots ,x_{N},t)=\int _{u_{1}=-\infty }^{\infty }\cdots \int _{u_{N}=-\infty }^{\infty }f_{C}(x_{1}-u_{1},\cdots ,x_{N}-u_{N},t)G(u_{1},t)\,du_{1}\cdots G(u_{N},t)\,du_{N},} where G ( x , t ) = 1 2 π t e − x 2 2 t {\displaystyle G(x,t)={\frac {1}{\sqrt {2\pi t}}}e^{-{\frac {x^{2}}{2t}}}} and the standard deviation of the Gaussian σ is related to the scale parameter t according to t = σ2. Separability will be assumed in all that follows, even when the kernel is not exactly Gaussian, since separation of the dimensions is the most practical way to implement multidimensional smoothing, especially at larger scales. Therefore, the rest of the article focuses on the one-dimensional case. == The sampled Gaussian kernel == When implementing the one-dimensional smoothing step in practice, the presumably simplest approach is to convolve the discrete signal fD with a sampled Gaussian kernel: L ( x , t ) = ∑ n = − ∞ ∞ f ( x − n ) G ( n , t ) {\displaystyle L(x,t)=\sum _{n=-\infty }^{\infty }f(x-n)\,G(n,t)} where G ( n , t ) = 1 2 π t e − n 2 2 t {\displaystyle G(n,t)={\frac {1}{\sqrt {2\pi t}}}e^{-{\frac {n^{2}}{2t}}}} (with t = σ2) which in turn is truncated at the ends to give a filter with finite impulse response L ( x , t ) = ∑ n = − M M f ( x − n ) G ( n , t ) {\displaystyle L(x,t)=\sum _{n=-M}^{M}f(x-n)\,G(n,t)} for M chosen sufficiently large (see error function) such that 2 ∫ M ∞ G ( u , t ) d u = 2 ∫ M t ∞ G ( v , 1 ) d v < ε . {\displaystyle 2\int _{M}^{\infty }G(u,t)\,du=2\int _{\frac {M}{\sqrt {t}}}^{\infty }G(v,1)\,dv<\varepsilon .} A common choice is to set M to a constant C times the standard deviation of the Gaussian kernel M = C σ + 1 = C t + 1 {\displaystyle M=C\sigma +1=C{\sqrt {t}}+1} where C is often chosen somewhere between 3 and 6. Using the sampled Gaussian kernel can, however, lead to implementation problems, in particular when computing higher-order derivatives at finer scales by applying sampled derivatives of Gaussian kernels. When accuracy and robustness are primary design criteria, alternative implementation approaches should therefore be considered. For small values of ε (10−6 to 10−8) the errors introduced by truncating the Gaussian are usually negligible. For larger values of ε, however, there are many better alternatives to a rectangular window function. For example, for a given number of points, a Hamming window, Blackman window, or Kaiser window will do less damage to the spectral and other properties of the Gaussian than a simple truncation will. Notwithstanding this, since the Gaussian kernel decreases rapidly at the tails, the main recommendation is still to use a sufficiently small value of ε such that the truncation effects are no longer important. == The discrete Gaussian kernel == A more refined approach is to convolve the original signal with the discrete Gaussian kernel T(n, t) L ( x , t ) = ∑ n = − ∞ ∞ f ( x − n ) T ( n , t ) {\displaystyle L(x,t)=\sum _{n=-\infty }^{\infty }f(x-n)\,T(n,t)} where T ( n , t ) = e − t I n ( t ) {\displaystyle T(n,t)=e^{-t}I_{n}(t)} and I n ( t ) {\displaystyle I_{n}(t)} denotes the modified Bessel functions of integer order, n. This is the discrete counterpart of the continuous Gaussian in that it is the solution to the discrete diffusion equation (discrete space, continuous time), just as the continuous Gaussian is the solution to the continuous diffusion equation. This filter can be truncated in the spatial domain as for the sampled Gaussian L ( x , t ) = ∑ n = − M M f ( x − n ) T ( n , t ) {\displaystyle L(x,t)=\sum _{n=-M}^{M}f(x-n)\,T(n,t)} or can be implemented in the Fourier domain using a closed-form expression for its discrete-time Fourier transform: T ^ ( θ , t ) = ∑ n = − ∞ ∞ T ( n , t ) e − i θ n = e t ( cos θ − 1 ) . {\displaystyle {\widehat {T}}(\theta ,t)=\sum _{n=-\infty }^{\infty }T(n,t)\,e^{-i\theta n}=e^{t(\cos \theta -1)}.} With this frequency-domain approach, the scale-space properties transfer exactly to the discrete domain, or with excellent approximation using periodic extension and a suitably long discrete Fourier transform to approximate the discrete-time Fourier transform of the signal being smoothed. Moreover, higher-order derivative approximations can be computed in a straightforward manner (and preserving scale-space properties) by applying small support central difference operators to the discrete scale space representation. As with the sampled Gaussian, a plain truncation of the infinite impulse response will in most cases be a sufficient approximation for small values of ε, while for larger values of ε it is better to use either a decomposition of the discrete Gaussian into a cascade of generalized binomial filters or alternatively to construct a finite approximate kernel by multiplying by a window function. If ε has been chosen too large such that effects of the truncation error begin to appear (for example as spurious extrema or spurious responses to higher-order derivative operators), then the options are to decrease the value of ε such that a larger finite kernel is used, with cutoff where the support is very small, or to use a tapered window. == Recursive filters == Since computational efficiency is often important, low-order recursive filters are often used for scale-space smoothing. For example, Young and van Vliet use a third-order recursive filter with one real pole and a pair of complex poles, applied forward and backward to make a sixth-order symmetric approximation to the Gaussian with low computational complexity for any smoothing scale. By relaxing a few of the axioms, Lindeberg concluded that good smoothing filters would be "normalized Pólya frequency sequences", a family of discrete kernels that includes all filters with real poles at 0 < Z < 1 and/or Z > 1, as well as with real zeros at Z < 0. For symmetry, which leads to approximate directional homogeneity, these filters must be further restricted to pairs of poles and zeros that lead to zero-phase filters. To match the transfer function curvature at zero frequency of the discrete Gaussian, which ensures an approximate semi-group property of additive t, two poles at Z = 1 + 2 t − ( 1 + 2 t ) 2 − 1 {\displaystyle
Text Retrieval Conference
The Text REtrieval Conference (TREC) is an ongoing series of workshops focusing on a list of different information retrieval (IR) research areas, or tracks. It is co-sponsored by the National Institute of Standards and Technology (NIST) and the Intelligence Advanced Research Projects Activity (part of the office of the Director of National Intelligence), and began in 1992 as part of the TIPSTER Text program. Its purpose is to support and encourage research within the information retrieval community by providing the infrastructure necessary for large-scale evaluation of text retrieval methodologies and to increase the speed of lab-to-product transfer of technology. TREC's evaluation protocols have improved many search technologies. A 2010 study estimated that "without TREC, U.S. Internet users would have spent up to 3.15 billion additional hours using web search engines between 1999 and 2009." Hal Varian the Chief Economist at Google wrote that "The TREC data revitalized research on information retrieval. Having a standard, widely available, and carefully constructed set of data laid the groundwork for further innovation in this field." Each track has a challenge wherein NIST provides participating groups with data sets and test problems. Depending on track, test problems might be questions, topics, or target extractable features. Uniform scoring is performed so the systems can be fairly evaluated. After evaluation of the results, a workshop provides a place for participants to collect together thoughts and ideas and present current and future research work.Text Retrieval Conference started in 1992, funded by DARPA (US Defense Advanced Research Project) and run by NIST. Its purpose was to support research within the information retrieval community by providing the infrastructure necessary for large-scale evaluation of text retrieval methodologies. == Goals == Encourage retrieval search based on large text collections Increase communication among industry, academia, and government by creating an open forum for the exchange of research ideas Speed the transfer of technology from research labs into commercial products by demonstrating substantial improvements retrieval methodologies on real world problems To increase the availability of appropriate evaluation techniques for use by industry and academia including development of new evaluation techniques more applicable to current systems TREC is overseen by a program committee consisting of representatives from government, industry, and academia. For each TREC, NIST provide a set of documents and questions. Participants run their own retrieval system on the data and return to NIST a list of retrieved top-ranked documents. NIST pools the individual result judges the retrieved documents for correctness and evaluates the results. The TREC cycle ends with a workshop that is a forum for participants to share their experiences. == Relevance judgments in TREC == TREC defines relevance as: "If you were writing a report on the subject of the topic and would use the information contained in the document in the report, then the document is relevant." Most TREC retrieval tasks use binary relevance: a document is either relevant or not relevant. Some TREC tasks use graded relevance, capturing multiple degrees of relevance. Most TREC collections are too large to perform complete relevance assessment; for these collections it is impossible to calculate the absolute recall for each query. To decide which documents to assess, TREC usually uses a method call pooling. In this method, the top-ranked n documents from each contributing run are aggregated, and the resulting document set is judged completely. == Various TRECs == In 1992 TREC-1 was held at NIST. The first conference attracted 28 groups of researchers from academia and industry. It demonstrated a wide range of different approaches to the retrieval of text from large document collections .Finally TREC1 revealed the facts that automatic construction of queries from natural language query statements seems to work. Techniques based on natural language processing were no better no worse than those based on vector or probabilistic approach. TREC2 Took place in August 1993. 31 group of researchers participated in this. Two types of retrieval were examined. Retrieval using an ‘ad hoc’ query and retrieval using a ‘routing' query In TREC-3 a small group experiments worked with Spanish language collection and others dealt with interactive query formulation in multiple databases TREC-4 they made even shorter to investigate the problems with very short user statements TREC-5 includes both short and long versions of the topics with the goal of carrying out deeper investigation into which types of techniques work well on various lengths of topics In TREC-6 Three new tracks speech, cross language, high precision information retrieval were introduced. The goal of cross language information retrieval is to facilitate research on system that are able to retrieve relevant document regardless of language of the source document TREC-7 contained seven tracks out of which two were new Query track and very large corpus track. The goal of the query track was to create a large query collection TREC-8 contain seven tracks out of which two –question answering and web tracks were new. The objective of QA query is to explore the possibilities of providing answers to specific natural language queries TREC-9 Includes seven tracks In TREC-10 Video tracks introduced Video tracks design to promote research in content based retrieval from digital video In TREC-11 Novelty tracks introduced. The goal of novelty track is to investigate systems abilities to locate relevant and new information within the ranked set of documents returned by a traditional document retrieval system TREC-12 held in 2003 added three new tracks; Genome track, robust retrieval track, HARD (Highly Accurate Retrieval from Documents) == Tracks == === Current tracks === New tracks are added as new research needs are identified, this list is current for TREC 2018. CENTRE Track – Goal: run in parallel CLEF 2018, NTCIR-14, TREC 2018 to develop and tune an IR reproducibility evaluation protocol (new track for 2018). Common Core Track – Goal: an ad hoc search task over news documents. Complex Answer Retrieval (CAR) – Goal: to develop systems capable of answering complex information needs by collating information from an entire corpus. Incident Streams Track – Goal: to research technologies to automatically process social media streams during emergency situations (new track for TREC 2018). The News Track – Goal: partnership with The Washington Post to develop test collections in news environment (new for 2018). Precision Medicine Track – Goal: a specialization of the Clinical Decision Support track to focus on linking oncology patient data to clinical trials. Real-Time Summarization Track (RTS) – Goal: to explore techniques for real-time update summaries from social media streams. === Past tracks === Chemical Track – Goal: to develop and evaluate technology for large scale search in chemistry-related documents, including academic papers and patents, to better meet the needs of professional searchers, and specifically patent searchers and chemists. Clinical Decision Support Track – Goal: to investigate techniques for linking medical cases to information relevant for patient care Contextual Suggestion Track – Goal: to investigate search techniques for complex information needs that are highly dependent on context and user interests. Crowdsourcing Track – Goal: to provide a collaborative venue for exploring crowdsourcing methods both for evaluating search and for performing search tasks. Genomics Track – Goal: to study the retrieval of genomic data, not just gene sequences but also supporting documentation such as research papers, lab reports, etc. Last ran on TREC 2007. Dynamic Domain Track – Goal: to investigate domain-specific search algorithms that adapt to the dynamic information needs of professional users as they explore in complex domains. Enterprise Track – Goal: to study search over the data of an organization to complete some task. Last ran on TREC 2008. Entity Track – Goal: to perform entity-related search on Web data. These search tasks (such as finding entities and properties of entities) address common information needs that are not that well modeled as ad hoc document search. Cross-Language Track – Goal: to investigate the ability of retrieval systems to find documents topically regardless of source language. After 1999, this track spun off into CLEF. FedWeb Track – Goal: to select best resources to forward a query to, and merge the results so that most relevant are on the top. Federated Web Search Track – Goal: to investigate techniques for the selection and combination of search results from a large number of real on-line web search services. Filtering Track – Goal: to binarily decide retrieval of new
AppValley
AppValley is an independent American digital distribution service operated and trademarked by AppValley LLC. It serves as an alternative app store for the iOS mobile operating system, which allows users to download applications that are not available on the App Store, most commonly tweaked "++" apps, jailbreak apps, and apps including paid apps on the app store. == Legality == AppValley is among several services that violate enterprise developer certificates from Apple. The terms under which these are granted make clear that they are for companies who wish to distribute apps to their employees. AppValley uses these certificates to distribute software directly to non-employees, thereby bypassing the AppStore. AppValley's conduct had implications in U.S. sanctioned markets like Iran, Iraq, North Korea, Cuba, and Venezuela, which have all been subject to commercial sanctions. Among the software offered by AppValley and other services is pirated software, including paid apps on the app store and premium versions of Instagram, Spotify, Pokémon Go, and others. For instance, AppValley distributes an ad-free version of the music streaming app Spotify even on the free tier. == History == The website was founded in May 2017, releasing late that month with a very basic version of the app. There were less than 100 apps available for download at this time. On Jan 19, 2018, a new version dubbed AppValley 2.0 was released bringing dark mode, more categories, a search, and a much faster interface. On February 14, 2019, a Chinese partner "Jason Wu" allegedly took control of the main Twitter account and domain, causing the original AppValley developers to migrate to the domain app-valley.vip and the Twitter account handle @App_Valley_vip. As of September 2024, the app-valley.vip domain now redirects to appvalley.signulous.com. Today, AppValley continues to offer an alternative to Apple's App Store where app developers can publish their applications. == Features == AppValley is a mobile app installer which can also support iOS version that can be installed and downloaded on the mobile or the devices of the people who wish to get access to many different applications available. AppValley also contains apps that have been modified or tweaked for user preferences, and allows the user to by pass national restrictions on the use of apps, without having to resort to jailbreaking. As of June 2, 2020, there are over 1300 apps available for download.
History of natural language processing
The history of natural language processing describes the advances of natural language processing. There is some overlap with the history of machine translation, the history of speech recognition, and the history of artificial intelligence. == Early history == The history of machine translation dates back to the seventeenth century, when philosophers such as Leibniz and Descartes put forward proposals for codes which would relate words between languages. All of these proposals remained theoretical, and none resulted in the development of an actual machine. The first patents for "translating machines" were applied for in the mid-1930s. One proposal, by Georges Artsrouni, was simply an automatic bilingual dictionary using paper tape. The other proposal, by Peter Troyanskii, a Russian, was more detailed. Troyanskii’s proposal included both the bilingual dictionary and a method for dealing with grammatical roles between languages, based on Esperanto. == Logical period == In 1950, Alan Turing published his famous article "Computing Machinery and Intelligence" which proposed what is now called the Turing test as a criterion of intelligence. This criterion depends on the ability of a computer program to impersonate a human in a real-time written conversation with a human judge, sufficiently well that the judge is unable to distinguish reliably — on the basis of the conversational content alone — between the program and a real human. In 1957, Noam Chomsky’s Syntactic Structures revolutionized Linguistics with 'universal grammar', a rule-based system of syntactic structures. The Georgetown experiment in 1954 involved fully automatic translation of more than sixty Russian sentences into English. The authors claimed that within three or five years, machine translation would be a solved problem. However, real progress was much slower, and after the ALPAC report in 1966, which found that ten years long research had failed to fulfill the expectations, funding for machine translation was dramatically reduced. Little further research in machine translation was conducted until the late 1980s, when the first statistical machine translation systems were developed. Some notably successful NLP systems developed in the 1960s were SHRDLU, a natural language system working in restricted "blocks worlds" with restricted vocabularies. In 1969 Roger Schank introduced the conceptual dependency theory for natural language understanding. This model, partially influenced by the work of Sydney Lamb, was extensively used by Schank's students at Yale University, such as Robert Wilensky, Wendy Lehnert, and Janet Kolodner. In 1970, William A. Woods introduced the augmented transition network (ATN) to represent natural language input. Instead of phrase structure rules ATNs used an equivalent set of finite-state automata that were called recursively. ATNs and their more general format called "generalized ATNs" continued to be used for a number of years. During the 1970s many programmers began to write 'conceptual ontologies', which structured real-world information into computer-understandable data. Examples are MARGIE (Schank, 1975), SAM (Cullingford, 1978), PAM (Wilensky, 1978), TaleSpin (Meehan, 1976), QUALM (Lehnert, 1977), Politics (Carbonell, 1979), and Plot Units (Lehnert 1981). During this time, many chatterbots were written including PARRY, Racter, and Jabberwacky. == Statistical period == Up to the 1980s, most NLP systems were based on complex sets of hand-written rules. Starting in the late 1980s, however, there was a revolution in NLP with the introduction of machine learning algorithms for language processing. This was due both to the steady increase in computational power resulting from Moore's law and the gradual lessening of the dominance of Chomskyan theories of linguistics (e.g. transformational grammar), whose theoretical underpinnings discouraged the sort of corpus linguistics that underlies the machine-learning approach to language processing. Some of the earliest-used machine learning algorithms, such as decision trees, produced systems of hard if-then rules similar to existing hand-written rules. Increasingly, however, research has focused on statistical models, which make soft, probabilistic decisions based on attaching real-valued weights to the features making up the input data. The cache language models upon which many speech recognition systems now rely are examples of such statistical models. Such models are generally more robust when given unfamiliar input, especially input that contains errors (as is very common for real-world data), and produce more reliable results when integrated into a larger system comprising multiple subtasks. === Datasets === The emergence of statistical approaches was aided by both increase in computing power and the availability of large datasets. At that time, large multilingual corpora were starting to emerge. Notably, some were produced by the Parliament of Canada and the European Union as a result of laws calling for the translation of all governmental proceedings into all official languages of the corresponding systems of government. Many of the notable early successes occurred in the field of machine translation. In 1993, the IBM alignment models were used for statistical machine translation. Compared to previous machine translation systems, which were symbolic systems manually coded by computational linguists, these systems were statistical, which allowed them to automatically learn from large textual corpora. Though these systems do not work well in situations where only small corpora is available, so data-efficient methods continue to be an area of research and development. In 2001, a one-billion-word large text corpus, scraped from the Internet, referred to as "very very large" at the time, was used for word disambiguation. To take advantage of large, unlabelled datasets, algorithms were developed for unsupervised and self-supervised learning. Generally, this task is much more difficult than supervised learning, and typically produces less accurate results for a given amount of input data. However, there is an enormous amount of non-annotated data available (including, among other things, the entire content of the World Wide Web), which can often make up for the inferior results. == Neural period == Neural language models were developed in 1990s. In 1990, the Elman network, using a recurrent neural network, encoded each word in a training set as a vector, called a word embedding, and the whole vocabulary as a vector database, allowing it to perform such tasks as sequence-predictions that are beyond the power of a simple multilayer perceptron. A shortcoming of the static embeddings was that they didn't differentiate between multiple meanings of homonyms. Yoshua Bengio developed the first neural probabilistic language model in 2000. Novel algorithms, availability of larger datasets and higher processing power made possible training of larger and larger language models. Attention mechanism was introduced by Bahdanau et al. in 2014. This work laid the foundations for the famous "Attention Is All You Need" paper that introduced the Transformer architecture in 2017. The concept of large language model (LLM) emerged in late 2010s. LLM is a language model trained with self-supervised learning on vast amount of text. Earliest public LLMs had hundreds of millions of parameters, but this number quickly rose to billion and even trillions. In recent years, advancements in deep learning and large language models have significantly enhanced the capabilities of natural language processing, leading to widespread applications in areas such as healthcare, customer service, and content generation. == Software ==
Bring your own encryption
Bring your own encryption (BYOE), also known as bring your own key (BYOK), is a cloud computing security model that allows cloud service customers to use their own encryption software and manage their own encryption keys. == Overview == BYOE enables cloud service customers to utilize a virtual instance of their encryption software alongside their cloud-hosted business applications to encrypt their data. In this model, hosted business applications are configured to process all data through the encryption software. This software then writes the ciphertext version of the data to the cloud service provider's physical data store and decrypts ciphertext data upon retrieval requests. This approach provides enterprises with control over their keys and the ability to generate their own master key using internal hardware security modules (HSM), which are then transmitted to the cloud provider's HSM. When the data is no longer needed, such as when users discontinue the cloud service, the keys can be deleted, rendering the encrypted data permanently inaccessible. This practice is known as crypto-shredding. == Potential Advantages == Organizations can store data with unique encryption that only they can access. Multiple organizations can share the same hardware infrastructure via cloud services like Amazon Web Services (AWS) or Google Cloud while maintaining encryption to comply with regulations such as HIPAA. == Potential Challenges == Resource utilization may be higher compared to traditional encryption practices when multiple users share the same hardware and use their own encryption. Efforts to minimize resource utilization issues may potentially impact security benefits.
Retrieval-augmented generation
Retrieval-augmented generation (RAG) is a technique that enables large language models (LLMs) to retrieve and incorporate new information from external data sources. With RAG, LLMs first refer to a specified set of documents, then respond to user queries. These documents supplement information from the LLM's pre-existing training data. This allows LLMs to use domain-specific and/or updated information that is not available in the training data. For example, this enables LLM-based chatbots to access internal company data or generate responses based on authoritative sources. RAG improves LLMs by incorporating information retrieval before generating responses. Unlike LLMs that rely on static training data, RAG pulls relevant text from databases, uploaded documents, or web sources. According to Ars Technica, "RAG is a way of improving LLM performance, in essence by blending the LLM process with a web search or other document look-up process to help LLMs stick to the facts." This method helps reduce AI hallucinations, which have caused chatbots to describe policies that don't exist, or recommend nonexistent legal cases to lawyers that are looking for citations to support their arguments. RAG also reduces the need to retrain LLMs with new data, saving on computational and financial costs. Beyond efficiency gains, RAG also allows LLMs to include sources in their responses, so users can verify the cited sources. This provides greater transparency, as users can cross-check retrieved content to ensure accuracy and relevance. The term retrieval-augmented generation (RAG) was introduced in a 2020 paper that described combining a parametric language model with a non-parametric external memory accessed through retrieval at inference time. == RAG and LLM limitations == LLMs can provide incorrect information. For example, when Google first demonstrated its LLM tool "Google Bard" (later re-branded to Gemini), the LLM provided incorrect information about the James Webb Space Telescope. This error contributed to a $100 billion decline in Google's stock value. RAG is used to prevent these errors, but it does not solve all the problems. For example, LLMs can generate misinformation even when pulling from factually correct sources if they misinterpret the context. MIT Technology Review gives the example of an AI-generated response stating, "The United States has had one Muslim president, Barack Hussein Obama." The model retrieved this from an academic book rhetorically titled Barack Hussein Obama: America's First Muslim President? The LLM did not "know" or "understand" the context of the title, generating a false statement. LLMs with RAG are programmed to prioritize new information. This technique has been called "prompt stuffing." Without prompt stuffing, the LLM's input is generated by a user; with prompt stuffing, additional relevant context is added to this input to guide the model's response. This approach provides the LLM with key information early in the prompt, encouraging it to prioritize the supplied data over pre-existing training knowledge. == Process == Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating an information-retrieval mechanism that allows models to access and utilize additional data beyond their original training set. Ars Technica notes that "when new information becomes available, rather than having to retrain the model, all that's needed is to augment the model's external knowledge base with the updated information" ("augmentation"). IBM states that "in the generative phase, the LLM draws from the augmented prompt and its internal representation of its training data to synthesize" an answer. === RAG key stages === Typically, the data to be referenced is converted into LLM embeddings, numerical representations in the form of a large vector space. RAG can be used on unstructured (usually text), semi-structured, or structured data (for example knowledge graphs). These embeddings are then stored in a vector database to allow for document retrieval. Given a user query, a document retriever is first called to select the most relevant documents that will be used to augment the query. This comparison can be done using a variety of methods, which depend in part on the type of indexing used. The model feeds this relevant retrieved information into the LLM via prompt engineering of the user's original query. Newer implementations (as of 2023) can also incorporate specific augmentation modules with abilities such as expanding queries into multiple domains and using memory and self-improvement to learn from previous retrievals. Finally, the LLM can generate output based on both the query and the retrieved documents. Some models incorporate extra steps to improve output, such as the re-ranking of retrieved information, context selection, and fine-tuning. == Applications == Retrieval-augmented generation is used in applications where generated responses need to be grounded in external or frequently updated information. Commonly cited use cases include search engines, question-answering systems, customer support chatbots, enterprise knowledge assistants, content generation, recommendation systems, retail and e-commerce, and industrial or manufacturing workflows. In healthcare, RAG has been studied as a way to ground large language model outputs in external medical knowledge sources, although reviews have noted continuing challenges around evaluation, ethics, and clinical reliability. == Improvements == Improvements to the basic process above can be applied at different stages in the RAG flow. === Encoder === These methods focus on the encoding of text as either dense or sparse vectors. Sparse vectors, which encode the identity of a word, are typically dictionary-length and contain mostly zeros. Dense vectors, which encode meaning, are more compact and contain fewer zeros. Various enhancements can improve the way similarities are calculated in the vector stores (databases). Performance improves by optimizing how vector similarities are calculated. Dot products enhance similarity scoring, while approximate nearest neighbor (ANN) searches improve retrieval efficiency over K-nearest neighbors (KNN) searches. Accuracy may be improved with Late Interactions, which allow the system to compare words more precisely after retrieval. This helps refine document ranking and improve search relevance. Hybrid vector approaches may be used to combine dense vector representations with sparse one-hot vectors, taking advantage of the computational efficiency of sparse dot products over dense vector operations. Other retrieval techniques focus on improving accuracy by refining how documents are selected. Some retrieval methods combine sparse representations, such as SPLADE, with query expansion strategies to improve search accuracy and recall. === Retriever-centric methods === These methods aim to enhance the quality of document retrieval in vector databases: Pre-training the retriever using the Inverse Cloze Task (ICT), a technique that helps the model learn retrieval patterns by predicting masked text within documents. Supervised retriever optimization aligns retrieval probabilities with the generator model's likelihood distribution. This involves retrieving the top-k vectors for a given prompt, scoring the generated response's perplexity, and minimizing KL divergence between the retriever's selections and the model's likelihoods to refine retrieval. Reranking techniques can refine retriever performance by prioritizing the most relevant retrieved documents during training. === Language model === By redesigning the language model with the retriever in mind, a 25-time smaller network can get comparable perplexity as its much larger counterparts. Because it is trained from scratch, this method (Retro) incurs the high cost of training runs that the original RAG scheme avoided. The hypothesis is that by giving domain knowledge during training, Retro needs less focus on the domain and can devote its smaller weight resources only to language semantics. The redesigned language model is shown here. It has been reported that Retro is not reproducible, so modifications were made to make it so. The more reproducible version is called Retro++ and includes in-context RAG. === Chunking === Chunking involves various strategies for breaking up the data into vectors so the retriever can find details in it. Three types of chunking strategies are: Fixed length with overlap. This is fast and easy. Overlapping consecutive chunks helps to maintain semantic context across chunks. Syntax-based chunks can break the document up into sentences. Libraries such as spaCy or NLTK can also help. File format-based chunking. Certain file types have natural chunks built in, and it's best to respect them. For example, code files are best chunked and vectorized as whole functions or classes. HTML files should leave