# Research & Analysis

> Source: https://aiwiki.ai/wiki/research_analysis
> Updated: 2026-06-28
> Categories: AI Research, AI Tools & Products, Data Science
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**Research & Analysis** in the AI sense refers to the use of [artificial intelligence](/wiki/artificial_intelligence), especially [large language models](/wiki/large_language_model) paired with [retrieval-augmented generation](/wiki/retrieval_augmented_generation), to find information, read sources, summarise documents, explore datasets, and write analytic deliverables grounded in citable evidence. The category spans general answer engines (Perplexity, ChatGPT Search), scientific literature assistants (Elicit, Consensus), document question-answering tools (NotebookLM), data-analysis copilots (ChatGPT Code Interpreter, Julius AI), and vertical products for legal, financial, and patent work (Harvey, Westlaw Precision, AlphaSense, PatSnap). The defining pattern across all of them is the [deep research](/wiki/deep_research) agent: a model that plans a query, browses or retrieves sources, then synthesises a referenced report instead of returning a list of links.

This page is a gateway to the tools and concepts in that space. It covers general purpose answer engines, scientific literature assistants, document question answering systems, data analytics copilots, spreadsheet copilots, business intelligence assistants, survey and user research platforms, and vertical products for legal, financial, and patent work.

The shift from traditional keyword search to retrieval-augmented language models began in 2020 and accelerated after the launch of [ChatGPT](/wiki/chatgpt) in November 2022. By 2025 the largest professional information vendors, Thomson Reuters, RELX (LexisNexis), Bloomberg, and S&P Global, had each shipped generative interfaces over their proprietary archives, and a wave of startups had built point solutions for systematic reviews, due diligence, financial modelling, and patent search. Most of these systems combine a hosted [LLM](/wiki/llm) with [retrieval-augmented generation](/wiki/retrieval_augmented_generation) so that answers are grounded in citable source documents rather than the model's parametric memory.

## Research & Analysis Custom GPTs

*See also: [Academic Research ChatGPT Plugins](/wiki/academic_research_chatgpt_plugins), [Search Engine ChatGPT Plugins](/wiki/search_engine_chatgpt_plugins), and [Data Visualization ChatGPT Plugins](/wiki/data_visualization_chatgpt_plugins).*

## What are general AI research assistants?

General research assistants are answer engines that combine a web crawl with a generative model. They differ from a classic [search engine](/wiki/search_engine) by writing a synthesised response with inline footnotes, instead of returning a list of links.

[Perplexity AI](/wiki/perplexity_ai), founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski, is the most widely used standalone product in this category. Perplexity's free tier serves a basic web answer, while Perplexity Pro routes through frontier models and adds longer follow-up threads. The company raised about $200 million in September 2025 at a $20 billion post-money valuation, with backers including Nvidia, Databricks, Jeff Bezos, and Tobias Lutke.[1][2] It was the third raise of 2025 after a $500 million Accel-led round at $14 billion in June and a $100 million extension at $18 billion in July, lifting total funding to roughly $1.5 billion.[2] Perplexity reports more than 30 million active users running over 780 million queries a month, with annual recurring revenue approaching $200 million.[2] A separate Sonar API exposes the underlying retrieval layer to developers.

ChatGPT Search is the search surface inside [ChatGPT](/wiki/chatgpt). [OpenAI](/wiki/openai) opened it to free users in late 2024 and added it to the default model behaviour in 2025. ChatGPT Search reads pages in real time, cites publishers, and routes long requests to the Deep Research mode that produces multi-page reports. Deep Research is benchmarked separately on systems such as [Deep Research Bench](/wiki/deep_research_bench) and [DeepResearch Bench](/wiki/deepresearch_bench).

[Claude](/wiki/claude) from [Anthropic](/wiki/anthropic) added native web search inside Claude.ai in 2025, with citations rendered alongside each paragraph. The Computer Use and tool-use APIs let third parties build their own research front ends on top of the same Claude models. [Gemini](/wiki/gemini) from [Google](/wiki/google) ships a Deep Research mode that runs longer multi-step browsing sessions and writes a report; the same mode is exposed through the Gemini Advanced subscription and inside Google Workspace.

[You.com](/wiki/you_com) launched in late 2021 and pivoted to generative answers in 2023 with You Chat. Phind, a Y Combinator company, focuses on developer questions; it ranks Stack Overflow, GitHub README files, and official docs ahead of general web pages and adds a code interpreter that runs Python in the browser. Brave Search AI, run by Brave Software, layers an LLM summary over Brave's independent web index and emphasises private browsing.

| Tool | Vendor | Model strategy | Differentiator |
|------|--------|----------------|----------------|
| [Perplexity](/wiki/perplexity) | Perplexity AI | Routes across GPT, Claude, and in-house Sonar | Standalone consumer product, Sonar API |
| ChatGPT Search | [OpenAI](/wiki/openai) | GPT family with Deep Research | Built into the chat product, large free tier |
| [Claude](/wiki/claude) with Search | [Anthropic](/wiki/anthropic) | Claude Opus, Sonnet, Haiku | Long context window, paragraph level citations |
| Gemini Deep Research | [Google](/wiki/google) | Gemini family | Tied to Workspace and Google Search index |
| [You.com](/wiki/you_com) | You.com Inc. | Multi-model router | Modes for research, writing, code |
| Phind | Phind Inc. | Custom and frontier models | Developer focus, in-browser code execution |
| Brave Search AI | Brave Software | Mixed | Independent web index, no third party trackers |

## Which AI tools handle academic and scientific literature?

General assistants struggle with peer reviewed sources because most journal articles sit behind paywalls and are not in the open web crawl. A separate group of products has been built specifically for scientific literature.

[Elicit](https://elicit.com), incubated at the non-profit research lab Ought and now an independent public benefit corporation, launched in 2023 with co-founders Andreas Stuhlmuller and Jungwon Byun. Elicit pulls publications from [Semantic Scholar](https://www.semanticscholar.org) and extracts study population, intervention, and outcome fields from each paper into a comparison matrix. The 2025 release added a structured systematic review workflow that runs an extraction template across hundreds of papers in parallel.

[Consensus](/wiki/consensus_gpt), founded in 2021 by Christian Salem and Eric Olson, indexes more than 220 million papers and answers natural language questions with a Consensus Meter that summarises agreement across the underlying literature.[6] Its Scholar Agent, released in 2025, is built on GPT-5 and the OpenAI Responses API, and splits the work across four specialised sub-agents (Planning, Search, Reading, and Analysis) to improve accuracy and reduce hallucination.[7] OpenAI reports that over the prior year Consensus grew to more than 8 million researchers and increased revenue 8x.[7]

Semantic Scholar is the underlying corpus that powers many of the products on this page. It is a free service from the Allen Institute for AI (AI2) and has indexed roughly 220 million papers as of 2026.[25] Its open API and the related Semantic Scholar Open Research Corpus (S2ORC) are widely used to build downstream tools.

Scite.ai introduces the concept of Smart Citations: it labels each citation as supporting, contrasting, or merely mentioning the cited paper, so that a reader can see at a glance whether a finding has been replicated. The Scite Assistant added an LLM front end in 2024 that answers questions and grounds each claim in a list of supporting and contrasting citations.

Connected Papers and ResearchRabbit visualise citation networks rather than answer questions in prose. Connected Papers builds a force-directed graph around a seed paper, useful for understanding the cluster of work most closely related to a single article. ResearchRabbit, acquired by Litmaps in 2025, supports ongoing literature collections that grow as new papers appear; it was re-released in November 2025 on a freemium plan.

| Tool | Vendor | Corpus | Primary use |
|------|--------|--------|-------------|
| Elicit | Elicit Inc. (formerly Ought) | Semantic Scholar | Systematic reviews, data extraction |
| [Consensus](/wiki/consensus_gpt) | Consensus | 220M papers | Yes/no questions with Consensus Meter |
| Semantic Scholar | Allen Institute for AI | 220M papers | Free academic search and API |
| Scite.ai | Research Solutions | Cited references | Smart Citations and Scite Assistant |
| Connected Papers | Connected Papers Ltd. | Semantic Scholar | Single-seed citation graphs |
| ResearchRabbit | Litmaps (acquired 2025) | Semantic Scholar, PubMed | Ongoing reading lists, alerts |

## How do AI tools answer questions over your own documents?

A second cluster of tools answers questions over documents that the user provides, rather than the open web. The pattern is sometimes called retrieval over private corpora or RAG over uploads.

[NotebookLM](/wiki/google_notebooklm) is Google's flagship product in this category. It launched as Project Tailwind at Google I/O in May 2023, powered by PaLM 2, and was rebranded to NotebookLM later that year; Google removed its experimental label on 17 October 2024.[3] Each notebook accepts up to a few hundred sources, including PDFs, Google Docs, web URLs, and YouTube transcripts; the model answers only from those sources and includes inline citations to the underlying passages. The Audio Overview feature, released in September 2024, generates a two-host podcast summary of the notebook and went viral in late 2024, and Google added interactive audio (the user can break in and ask the hosts a question) in December 2024.[4] NotebookLM Plus, launched in December 2024 for Workspace and Gemini Advanced subscribers, raised the document and notebook limits and added shareable team notebooks.[4] Video Overviews, slide decks, and structured Data Tables outputs followed in 2025.

ChatGPT and [Claude](/wiki/claude_ai) both support direct file uploads with question answering against the contents. ChatGPT exposes this in the standard chat window and through Custom GPTs, where a knowledge base of up to 20 files can be attached to a tailored assistant. Claude's Projects feature stores up to 200,000 tokens of reference material and reuses it across every conversation in the project.

Humata, launched in 2022, is a focused PDF question answering product used in regulated industries; it advertises GDPR compliant hosting and a team workspace that retains source citations. Scholarcy summarises long PDFs into structured flashcards that highlight study purpose, sample size, methods, and findings; it is widely used in graduate programmes and research libraries.

## How do AI data-analysis tools work?

For work that involves running calculations on a dataset, the dominant tools are LLM-driven Python and SQL environments rather than chat surfaces.

ChatGPT's Code Interpreter, also marketed as Advanced Data Analysis and as the Python tool inside ChatGPT, runs sandboxed Python with NumPy, pandas, scikit-learn, and matplotlib. Users upload a CSV or Excel file and ask questions in plain language; the model writes code, runs it, and shows both the chart and the underlying script. The same pattern appears in Claude's Analysis tool and in Gemini's Python execution environment.

[Julius AI](/wiki/julius_ai) is a standalone product that competes directly with Code Interpreter for analytics workflows. Julius accepts CSV, Excel, PDF, and live database connections to Snowflake, BigQuery, Databricks, and Postgres; its Pro plan advertises support for files up to 32 GB, much larger than the 512 MB cap on ChatGPT uploads.[27] The product writes Python and R, builds visualisations, runs predictive models, retains state across a session, and produces formatted memos for stakeholders.

Hex is a notebook environment for data teams that combines SQL, Python, and a no-code report builder in a single canvas. Its Notebook Agent, released in 2025, is built on Claude Sonnet 4 and chains together SQL queries, transformations, and visualisation cells to answer a single business question.[23] Hex publishes case studies with Anthropic, Ramp, Figma, and others.

Polymer is a self-serve business intelligence tool that turns a Google Sheet, CSV, or warehouse table into an interactive dashboard with no setup. It bundles its own AI Assistant that suggests visualisations and writes commentary on a chart's findings.

## Spreadsheet copilots

Spreadsheets remain the working surface for most analysts, and the major office suites have integrated AI directly into the cell grid.

[Microsoft Copilot](/wiki/microsoft_copilot) for Excel, part of Microsoft 365 Copilot, generates formulas, formats tables, runs what-if analyses, and writes summaries from a worksheet. It can produce charts from a natural language prompt and surface anomalies in long tables. The Excel implementation runs alongside a Python in Excel feature that lets users execute Python code on data inside a workbook.

Gemini in Google Sheets adds a side panel that drafts formulas, generates charts, and answers questions about the active sheet. It is included with the Gemini Business and Enterprise tiers of Google Workspace. The Sheets implementation reads the cell grid as structured context, so the model can refer to specific named ranges in its replies.

Equals is a query-driven spreadsheet built around natural language SQL: every cell can pull data live from a connected database without writing a query by hand. Rows is a similar product oriented to business operations teams, with prebuilt connectors for Stripe, HubSpot, Google Analytics, and Shopify. Both products advertise an LLM that converts a plain English question into the structured query under the hood.

| Tool | Vendor | Surface | Notable feature |
|------|--------|---------|-----------------|
| Excel Copilot | [Microsoft](/wiki/microsoft) | Microsoft 365 | Formulas, charts, anomaly detection, Python in Excel |
| Gemini in Sheets | [Google](/wiki/google) | Google Workspace | Side panel, named range awareness |
| Equals | Equals Inc. | Standalone web sheet | Live database queries from cells |
| Rows | Rows GmbH | Standalone web sheet | Prebuilt SaaS connectors |

## Business intelligence and data visualisation

The established BI vendors have shipped generative layers that sit on top of an existing semantic model.

Tableau Pulse from Salesforce announced in May 2023 and reached general availability in February 2024.[21] Pulse watches a small set of business metrics defined by an analyst, then writes plain English summaries of how each metric is moving and which dimensions are driving the change. Pulse for Salesforce, released in March 2025, embeds the same insights inside Sales Cloud. Pulse is included with all Tableau Cloud editions.

Power BI Copilot, part of Microsoft Fabric, was generally available across Power BI from 2024.[22] Inside a Power BI report it can author a new page from a prompt, write Data Analysis Expressions (DAX) for measures, generate narrative summaries, and answer questions across the full report rather than a single page. A standalone Copilot experience added in 2025 lets users ask questions across every report and semantic model they have access to in their tenant; a Fabric data agent can be invoked from the same surface for questions over an entire warehouse.

Looker (from [Google](/wiki/google) Cloud), Qlik, ThoughtSpot, and Domo have all added comparable Copilot features, generally on the same RAG over a semantic model pattern. The point is that the LLM is constrained to the dimensions and measures the analyst has defined, which keeps the answers grounded and prevents the model from inventing fields.

## Survey and user research

Product teams use AI to read and cluster qualitative interviews and survey responses.

Dovetail is a research repository that automatically tags transcripts, clusters similar quotes into themes, and generates an executive summary at the project level. It has integrations with Zoom, Microsoft Teams, and Otter for ingesting recorded sessions.

Maze is a usability testing platform that uses AI to draft moderator scripts, score open ended responses, and surface usability friction across a study. The 2025 release added an AI Interviewer that conducts and transcribes unmoderated voice interviews at scale. Both tools are positioned for product, design, and UX teams rather than academic researchers.

## How is AI used for legal research?

Legal information is one of the most heavily invested vertical applications, because the work is high value, the source material is well structured, and law firms have a long history of paying for premium databases.

[Harvey](/wiki/harvey_ai) was incubated by the OpenAI Startup Fund in late 2022 and is the highest profile pure-play legal AI startup. On 25 March 2026 the company raised $200 million in a round co-led by GIC and Sequoia Capital at an $11 billion valuation, lifting total funding past $1 billion and roughly 3.5x-ing its valuation within a year.[14][15] Harvey reports that more than 100,000 lawyers run their most critical work on the platform, on which customers had built over 25,000 custom agents.[15] Harvey is built for law firm workflows: contract review, deposition prep, due diligence, and litigation drafting. Its agents pull from firm document management systems and from public legal corpora.

Casetext launched CoCounsel in March 2023 as one of the first GPT-4 based legal assistants. Thomson Reuters acquired Casetext in August 2023 for $650 million in cash.[16] CoCounsel 2.0, released in 2024, runs across multiple frontier models and feeds Westlaw content to the LLMs. Westlaw Precision, the flagship Thomson Reuters research product, uses RAG over the Westlaw classification system and editorial archive (more than 150 years of head notes and key numbers). Thomson Reuters reported more than 1.5 million AI Assisted Research searches across about 6,000 customers in the first year. In August 2025 Thomson Reuters launched CoCounsel Legal with Deep Research, an agentic capability that produces multi-section memos with arguments on both sides of an issue.[17]

Lexis+ AI, launched in 2023 by RELX subsidiary LexisNexis, is the equivalent product on the LexisNexis platform. The next generation, Lexis+ with Protege, was announced in May 2026 and adds workflow automation, agentic task completion, and routing across LLMs from Anthropic, Google, and OpenAI.[18] The platform is built on LexisNexis editorial content and Shepard's citator intelligence; customer inputs are not used to train any underlying LLM.

A second wave of niche legal AI tools has emerged for specific tasks: Eve and EvenUp for plaintiffs' personal injury work, Spellbook for transactional drafting, and LegalGPT-style assistants on top of base ChatGPT (see [LegalGPT](/wiki/legalgpt) and [Legal ChatGPT Plugins](/wiki/legal_chatgpt_plugins)).

| Tool | Vendor | Status | Use case |
|------|--------|--------|----------|
| [Harvey](/wiki/harvey_ai) | Harvey Inc. | $11B valuation, March 2026 | Big law workflows, due diligence |
| Casetext CoCounsel | Thomson Reuters | Acquired August 2023 | Document review, depositions |
| Westlaw Precision AI | Thomson Reuters | RAG over Westlaw archive | Statutory and case law research |
| Lexis+ with Protege | RELX (LexisNexis) | Launched May 2026 | Drafting, research, agentic workflows |

## How is AI used for financial research?

Financial research has its own set of vertical AI products focused on filings, transcripts, and structured market data.

Bloomberg has integrated AI throughout its terminal. The company published [BloombergGPT](https://arxiv.org/abs/2303.17564) in March 2023, a 50 billion parameter model trained on 363 billion tokens of Bloomberg's proprietary financial corpus (called FinPile) plus 345 billion tokens of general purpose data.[12][13] The model powers natural language search, sentiment analysis on news, and conversion of plain English questions into Bloomberg Query Language (BQL). The terminal's later AI Assistant features (transcript summarisation, document Q&A, earnings comparison) sit on top of the same internal stack.

[Hebbia](/wiki/hebbia), founded in 2020 by former Stanford PhD student George Sivulka, raised $130 million in a Series B led by Andreessen Horowitz in July 2024 at a $700 million valuation.[8][9] Hebbia's Matrix product is a large-scale document RAG system used by hedge funds and private equity firms for due diligence. The company reports that its platform indexes more than 1 billion pages and that roughly a third of the largest global asset managers by assets under management are customers, a base it has grown partly through a September 2025 partnership that brings FactSet market and estimates data into the platform.[8]

AlphaSense, founded in 2011 by Jack Kokko and Raj Neervannan, started as a search engine over financial filings and earnings call transcripts and added a generative summarisation layer in 2023. On 3 June 2026 the company raised $350 million at a $7.5 billion valuation (nearly double its prior $4 billion mark) and reported surpassing $600 million in annual recurring revenue; 88 percent of the S&P 100 are customers and its proprietary library spans more than 500 million business documents.[10][11] The 2024 acquisition of Tegus expanded its expert call transcript library. "This milestone reflects both the accelerating global adoption of our platform, and a broader shift in market intelligence: from fragmented information to end-to-end AI-driven workflows," said co-founder and chief executive Jack Kokko.[11]

Daloopa, founded in 2019 in New York by Thomas Li, Jeremy Huang, and Daniel Chen, automates financial model building. Its agents extract structured KPIs from filings, presentations, and supplements, then write the values directly into an analyst's Excel model with cell-level audit links back to the source.[19][20] The company reports coverage of 5,500+ tickers and serves both buy-side and sell-side firms.

See also [AI in finance](/wiki/ai_in_finance) and [Finance ChatGPT Plugins](/wiki/finance_chatgpt_plugins) for adjacent products.

| Tool | Vendor | Coverage | Differentiator |
|------|--------|----------|----------------|
| BloombergGPT and Bloomberg AI Assistant | Bloomberg L.P. | Bloomberg terminal data | 50B parameter domain model, BQL generation |
| [Hebbia](/wiki/hebbia) | Hebbia Inc. | Buy-side documents | Document agents at billion-page scale |
| AlphaSense | AlphaSense Inc. | 500M+ filings and transcripts | Search plus summary, 88% of the S&P 100 |
| Daloopa | Daloopa Inc. | 5,500+ tickers | Auto-builds Excel models with cell-level citations |

## Patent and intellectual property research

Patent search is a specialised research domain because the corpus is large (over 200 million documents from more than 180 patent offices), the language is technical and domain-specific, and the consequences of a missed prior art reference can run to many millions of dollars.

PatSnap is the dominant AI-driven patent platform, with more than 15,000 customers worldwide.[24] It indexes 202 million+ patents along with non-patent literature and chemical structure data, and translates documents from major filing languages into English. PatSnap shipped a domain-specific LLM in 2024 trained on its proprietary innovation data; its Novelty Search AI Agent uses RAG over claims and prior art to reduce model hallucination, and the company publishes a PatentBench benchmark on which it reports its agent outperforming general-purpose LLMs on novelty search.[24] The platform also integrates with FreedomToOperate and competitive landscape workflows.

Other vendors in the space include LexisNexis IP (PatentSight, TotalPatent One), Clarivate (Derwent Innovation), and Questel (Orbit Intelligence), each of which has shipped a generative answer feature on top of its existing patent corpus.

## What are the risks of AI research and analysis tools?

Research and analysis is one of the highest stakes domains for [LLM](/wiki/llm) deployment, because errors in a citation or a financial figure can survive into a published memo or filing. Several recurring failure modes are worth flagging.

Fabricated citations, sometimes called [hallucinations](/wiki/hallucination), are the most public risk. In the 2023 case Mata v. Avianca in the Southern District of New York, attorneys submitted a brief that cited six cases (Varghese, Shaboon, Petersen, Martinez, Durden, and Miller), none of which existed. The cases had been generated by ChatGPT and the lawyer had not checked them against a primary source. On 22 June 2023 Judge P. Kevin Castel fined the two attorneys and their firm $5,000, finding they had acted in "subjective bad faith" and describing one of the fabricated opinions as "gibberish."[26] The case is now a standard reference for why every legal AI vendor stresses retrieval grounding and citation links back to a verifiable document. Even high quality systems still occasionally invent a section number, swap two case names, or paraphrase a holding into something that the underlying source does not actually say. Cross-checking each citation against the primary source remains a basic professional duty.

Data confidentiality is the second concern, especially in finance and law. Sending a client document to a public chat surface can violate engagement letters, attorney-client privilege, GDPR, HIPAA, or institutional confidentiality policies. Vendors targeting professional users address this with enterprise tiers that contractually exclude inputs from training, host the model in a single-tenant or BYO-cloud configuration, and pass SOC 2 Type II and ISO 27001 audits. The Lexis+ with Protege launch announcement, for example, explicitly states that customer inputs are not used to train any LLM.[18] Buyers should confirm the relevant data processing addendum and check whether the vendor uses retention windows for prompt logs.

A third concern is reproducibility. Model outputs vary between runs and across model versions, so the same prompt can return different summaries on different days. For systematic reviews and regulated workflows, teams typically pin a model version, log every prompt and response, and treat the LLM as a draft tool rather than a final author. Many of the vertical products listed above ship audit logs and revision histories specifically to address this.

Finally, evaluation is hard. Standard benchmarks like [Deep Research Bench](/wiki/deep_research_bench), [DeepResearch Bench](/wiki/deepresearch_bench), and [LegalBench](/wiki/legalbench) provide a starting point, but real research work involves judgement calls that benchmarks do not capture. Most professional teams run their own internal accuracy reviews against a known good corpus before rolling a tool out broadly.

## ELI5: How does an AI research assistant work?

Imagine you ask a very fast librarian a question. Instead of just guessing the answer from memory (which is how a plain chatbot works and how it can make things up), the AI research assistant first runs off to the shelves, pulls down the actual books and articles that mention your topic, reads the relevant pages, and then writes you a short summary with little numbered notes pointing at exactly which book each fact came from. That extra step of going to the real sources first is called [retrieval-augmented generation](/wiki/retrieval_augmented_generation), and it is what separates a research tool you can trust from a chatbot that might invent a fact. A [deep research](/wiki/deep_research) agent goes one step further: it plans the whole project, keeps fetching more sources until it has enough, and hands back a multi-page report with citations, much like a research assistant who works for hours on their own and then shows you the finished memo.

## See also

- [Academic Research](/wiki/academic_research)
- [Academic Research ChatGPT Plugins](/wiki/academic_research_chatgpt_plugins)
- [Search Engine](/wiki/search_engine)
- [Search Engine ChatGPT Plugins](/wiki/search_engine_chatgpt_plugins)
- [Data Visualization](/wiki/data_visualization)
- [Data Analysis](/wiki/data_analysis)
- [Retrieval-Augmented Generation (RAG)](/wiki/retrieval_augmented_generation_rag)
- [Information Retrieval](/wiki/information_retrieval)
- [Semantic search](/wiki/semantic_search)
- [AI in finance](/wiki/ai_in_finance)
- [Legal](/wiki/legal)
- [LegalBench](/wiki/legalbench)
- [Hallucination](/wiki/hallucination)
- [Deep Research](/wiki/deep_research)

## References

1. "Perplexity AI." Wikipedia. Accessed June 2026.
2. TechCrunch. "Perplexity reportedly raised $200M at $20B valuation." 10 September 2025.
3. "NotebookLM." Wikipedia. Accessed June 2026.
4. Google. "NotebookLM gets a new look, audio interactivity and a premium version." The Keyword (blog.google), December 2024.
5. Ought. "Elicit." ought.org/elicit. Accessed June 2026.
6. Consensus. "About Us." consensus.app/home/about-us. Accessed June 2026.
7. OpenAI. "Consensus uses GPT-5 and the Responses API to complete weeks of research in minutes." openai.com, 2025.
8. "Hebbia." Wikipedia. Accessed June 2026.
9. TechCrunch. "AI startup Hebbia raised $130M at a $700M valuation on $13 million of profitable revenue." 9 July 2024.
10. AlphaSense. "What is AlphaSense?" Accessed June 2026.
11. AlphaSense. "AlphaSense Raises $350M at $7.5B Valuation, and Surpasses $600M in Annual Recurring Revenue." alpha-sense.com, 3 June 2026.
12. Bloomberg. "Introducing BloombergGPT." 30 March 2023.
13. Wu, S. et al. "BloombergGPT: A Large Language Model for Finance." arXiv:2303.17564, 2023.
14. CNBC. "Legal AI startup Harvey valued at $11 billion in funding round." 25 March 2026.
15. Harvey. "Harvey Raises Growth Round at $11 Billion Valuation Co-led by GIC and Sequoia." harvey.ai, 25 March 2026.
16. Thomson Reuters. "Thomson Reuters Acquires Legal AI firm Casetext for $650 Million." June 2023.
17. Thomson Reuters. "Thomson Reuters Launches CoCounsel Legal with Agentic AI and Deep Research Capabilities." August 2025.
18. LexisNexis. "LexisNexis Launches Next Evolution of Lexis+ with Protege." 7 May 2026.
19. Daloopa. "About Daloopa." Accessed June 2026.
20. TechCrunch. "Credit Suisse leads $20M Series A in data extraction startup Daloopa." 15 July 2021.
21. Salesforce. "Tableau Pulse Now Available." February 2024.
22. Microsoft Learn. "Copilot for Power BI overview." Accessed June 2026.
23. Hex Technologies. "Introducing the Notebook Agent." 2025.
24. PatSnap. "The AI-Native Platform for Global Innovation" and "PatentBench." patsnap.com. Accessed June 2026.
25. Allen Institute for AI. "Semantic Scholar." Accessed June 2026.
26. Mata v. Avianca, Inc., 678 F.Supp.3d 443, 22-cv-1461 (PKC) (S.D.N.Y. 22 June 2023).
27. Julius AI. "Julius vs ChatGPT" and pricing pages. julius.ai. Accessed June 2026.