Run your own version of Perplexity with a single Python program I wrote a single Python program to implement the search-extract-summarize flow, similar to AI search engines such as Perplexity. Code is here: github.com/pengfeng/ask.py The single python program also includes a full functional web UI using @Gradio ! A demo is running in @huggingface Space here: huggingface.co/spaces/LeetTo… Basically, given a query, the program will - search Google for the top 10 web pages - crawl and scape the pages for their text content - use Docling from @Ibm Deep Search Team (ds4sd.github.io/docling/) to convert any document into markdown - use @ChonkieAI (github.com/chonkie-ai/chonki…) to chunk the text content into chunks and save them into a @duckdb with both vectors and full text - perform a hybrid search in DuckDB using both vector and full text search with the query to find the top 10 matched chunks - use the top 10 chunks as the context to ask an LLM to generate the answer - output the answer with the references. And it even supports more advanced functions like - extract structured data from the web page - specify the target web site to limit the search - specify different language to output the results - limit or extend the answer to desired length - and it supports query local files as well Have fun and enjoy!

It actually reminds me of working with junior engineers: 1. you have to break down the task into reasonable blocks that can be verified and reverted; 2. the result usually works but will probably introduce some tech debt if you do not review carefully; 3. if they need to fix something, they usually just fix that one thing without realizing there are interactions with other components unless you tell them; 4. they are pretty bad at interface design which leads to very in-flexible and convoluted programs need a lot of refactoring. But (human) junior engineers usually learn pretty fast and improve on things you point out. Right now the agents can make the same mistakes many many times. And yes the productivity gains are huge even with all the current limitations. The potential is really unlimited.

Checking old pictures with @rangadi

Netherlands team taking their victory lap. What a game! #WC2014

Finally, Arena Fonte Nova! #wc2014

The previous SWE team was built on the premise of task composition, so a lot of time was spent in task decomposition, assigning different blocks to different people based on ability/velocity/specialization, trying to parallelizing the execution as much as possible and then combining the results to form the final results. There are several obvious inefficiencies in this process such as synchronized blocking and cross-human-interface information passing, that’s why senior engineers oftentimes prefer to do everything by themselves and great eng orgs try a lot of ways to remove those inefficiencies. Now with AI-enabled coding, the software development process will change a lot : 1) The granularity of the tasks can be much coarser, and thus less information exchange across the human interface; 2) the way to decompose the task will be more vertical, less horizontal, you don’t separate front-end from backend or designer from front-end anymore; 3) the expectations of the ability/velocity/specialization of every member in the team will be increased. The key function of every member is not generating code anymore, it will be tasks such as asking the right question to the AI coding agent and making the correct decision when the AI agent asks for a decision. So basically any SWE will need a broad understanding of the fields, know the best practice of different kinds of subtasks in the industry, can ask the right important questions, and make the right decision when facing different choices (otherwise you will be a liability in a team instead of an asset). Maybe it is not life-critical as the doctors and lawyers’ jobs but we will need a very different training process than the current CS undergrad curriculum.

Run your own open source Deep Research with DeepSeek-r1 Both OpenAI o1-pro model and Google Gemini 1.5-pro model now provide the "Deep Research" function that allows users to generate a research report based on a query. Our open source project LeetTools actually provides a similar tool that can work with any LLM model with text extract and summarize functions. We will use the DeepSeek model API from @FireworksAI_HQ as an example. Just a simple installation and one simple config file, you can run your own Deep Research! More interestingly, since we are using the R1 model, the output will show how the model is constructing the results and the thinking trace of each section as show in the figure.

Great discussion from @martin_casado and @stevesi about the new changes brought by DeepSeek and how the current AI trend compares to previous tech waves. The top takeaways from the discussion: - the growth trajectory of the current AI industry is more similar to Internet than the other tech innovations, and right now it is similar to pre-dotcom when the infras were being built. - the best opportunities are in the application layer to use right models to solve right problems; Deepseek proves that apps can now scale out with edge-size specialized models. - the moat will not be in the model size or ability, instead it will be in the application layer in terms of user experience, functionality, and distribution channel. There are many gems in the pod-cast such as - how to get to 7B endpoints - the switch war and the changing benchmarks - what is different this time around - why dull things like single sign-on and RBAC are important - why OpenAI and Anthropic will be fine Great listen, highly recommended. piped.video/watch?v=7lfTewdP…

"no extra abstraction" is actually super important: Low code platforms limit you to the design and creativity of the platform builders; AI coding let you work basically on the languages and all the libs directly without the "low-entropy keystrokes"; you are only limited by your own creativity and understanding of the whole stack.

It is crazy how things are moving these days. We can now run the deepseek-r1 model (the 1.5b version) with fully local AI search pipeline on a laptop using only 4GB of memory and no GPU is required (thanks to @ollama and @duckdb ). We have a full AI search demo using Ollama running on a laptop, and here is the thinking process output by deepseek-r1 when we ask the question "How does the Meme Coin work?", which is pretty good!

Yeah, I hope the cursor team is working on some thing along this line so that the agents can learn from this kind of feedback loops. Also I really like the new chat history export function so that I can summarize my own feedback patterns as well (still waiting to test it though).

Stamped

VLDB 2014 "SQL-on-Hadoop: Full Circle Back to Shared-Nothing Database Architectures" vldb.org/pvldb/vol7/p1295-fl…

Go terps! #FosterFarmsBowl

wow finally, looking good! can't believe it's been 10 years now:-)

3 lessons in database design from the team behind Twitter's Manhattan gigaom.com/2014/05/12/3-less… via @gigaom

"LLMs are so socially dumb" -> lol basically Internet is socially dumb.

Have been using @ChonkieAI since the first day they released, great API and great performance, well designed. Grats on the launch!

so proud of my little girl! facebook.com/dancekar/posts/…

@kurtosis0 "Review Request: removing baby" certainly got everyone's attention!

Introducing Twitter Data Grants blog.twitter.com/2014/introd…

Pain in the gas

I just 3-shotted a Reddit sentiment analysis pipeline using @cursor_ai and new GPT-5 from @OpenAI . The result is kind of nuts: it first searches on Google on a specified topic, get the top ones, then go to Reddit to retrieve the related posts, and then extract the sentiments from the posts. All the search data and results are saved to a database as well. It was not possible even a week ago. Why 3-shots? Because even the best product manager needs feedback from the real results, so the ability to evolve the program is actually more important than to one-shot stuff. The prompts and results are shown in screenshots below.

@luby @jco @posco where is the DTD? we need metadata @billgraham

This is a great summary and blueprint for "AI-native" or "Agent-native" apps: "Software becomes about reviewing the AI Agents’ work, creating task queues, seeing the status of work as it’s happening, being able to interrupt the work when necessary, checking audit trails, managing parallel threads, connecting work outputs from multiple agents, and more." So basically the infra work becomes more important now.

@kurtosis0 @ColorGenomics congratulations!

Using AI-assisted coding efficiently is very similar to being a good eng manager: 0. you have to establish good eng practices and guiding rules; 1. you need to understand the team members' abilities, strength, and weakness; 2. you must assign the tasks in precise and clear languages, preferably with good examples, and expected outcomes; 3. you can measure the team members' output quality and point out inefficiencies or inconsistent styles; 4. if the team member can't do the job, you have to be able to get it done yourself.

Perfect SWE mentality from Claude Code: not my code, not my problem.

BBC News - Google achieves AI 'breakthrough' by beating Go champion bbc.com/news/technology-3542…

I think we will see more AI-native or AI-friendly infra projects coming out. Relational is still efficient and great, but the usage pattern must change to fit the new era: SQL is a language and LLM will be better at it than us (maybe not right now). Supabase snaps $200M at $2B valuation to power the future of vibe coding: 3 things to know: - Developer-friendly and open-source by design - Perfect for the AI era and vibe coding - Strong traction with startups and enterprises alike techfundingnews.com/supabase… via @TFNBreakingNews

Running your own simplified version of Perplexity with one Python file: github.com/pengfeng/ask.py

@thealienlabs @cgst @mirceap nice!

So-Called “Twitter Tax Break” Netted City More In Mid-Market Tax Revenue tcrn.ch/1oRjUgs via @techcrunch

How to make your i18n automatic using Cursor Add a cursor rule under .cursor/rules/i18n.mdc: ``` ## i18n Our i18n files are under: frontend/src/i18n/{ja,en,zh}/{module_name}.js For string values used in the frontend, please use the following format: import i18n from '@/i18n' i18n.t('common.blankDocumentNotification') where `common` is the module name, and `blankDocumentNotification` is the key. ```

With all the sensationalism in the news and reports, we need some real human-in-the-loop authentic evaluations. Predefined benchmarks are just too easy to manipulate or overfit.

@squarecog truly the end of an era, tks for everything, and best wishes!

Damn o3 in Cursor is expensive!

@pankaj congratulations! Best luck with your new adventure back home.

Twitter wall

After Anthropic's Model Context Protocol (MCP) was introduced a few days ago, today we get a new agent framework from previous Android team: /dev/agents. So many frameworks .. lol Amazon: Amazon multi-agent-orchestrator OpenAI: Swarm Microsoft: AutoGen Facebook: LlamaStack Google: Vertex AI Agent Builder IBM: Bee Agent Framework CrewAI LangChain LlamaIndex n8n dify coze

Definitely an interesting read from Bloomberg. It has been almost 5 years since the original scaling law paper (arxiv.org/pdf/2001.08361), wondering if anyone will update the following figures in the paper. Since a lot of func gains actually came from things like MoE, RLHF, CoT, and agents, I think there is still a long way to go even if we just want to utilize the full potential of the current base models.

Since both OpenAI O1-pro and Google Gemini 1.5-pro have now provided the Deep Research function, I tried both with some real questions and want to share the results and some of our observations here. Here are the questions: 1. How will agentic AI and generative AI affect our non-tech jobs? (from reddit) 2. What is interviewing like now with everyone using AI? (from hackernews) 3. Help me research recent AI-powered marketing campaigns to benchmark for 2025 planning (this is a Gemini sample question) Their answers and our observations are listed below.

One problem AI coding assistant can't solve at least in the near future is to decide what to do, which is actually the most important task in a project, and in every day dev workflow. All the metrics citing how many lines the tools have been committing are kind of misleading. The real important code in any meaningful software project is less than 5% of the lines usually (I suspect even less in many cases). More lines of code brings side effects such as more bugs, more difficult to understand, more tests to maintain, and etc.

This is a really nice use case, maybe a little bit too advanced for most people. I use Cursor as a Finder replacement sometimes for the search and preview functions. If the file list view can be integrated (and the memory usage controlled a little bit better), it will be my default finder for sure. Also I can totally see photo search (or any local search function) implemented in this way.

“Much to the dismay of my department head, programs were never proven or formally verified. And so on.‘ lol at this one, with English being the new programming language it is impossible now!

Pretty cool demo from @AnthropicAI on the "Computer Use" model. One of the prompts used is show below: (Demo code here: github.com/anthropics/anthro…) I am wondering how can we scale this kind of approach to complex real world envs. Maybe we will have hundreds of "AI drivers/Adaptors" for each type of systems to do this kind of setup, like the audio/video drivers for each type of cards in the old days. SYSTEM_PROMPT = f"""<SYSTEM_CAPABILITY> * You are utilising an Ubuntu virtual machine using {platform.machine()} architecture with internet access. * You can feel free to install Ubuntu applications with your bash tool. Use curl instead of wget. * To open firefox, please just click on the firefox icon. Note, firefox-esr is what is installed on your system. * Using bash tool you can start GUI applications, but you need to set export DISPLAY=:1 and use a subshell. For example "(DISPLAY=:1 xterm &)". GUI apps run with bash tool will appear within your desktop environment, but they may take some time to appear. Take a screenshot to confirm it did. * When using your bash tool with commands that are expected to output very large quantities of text, redirect into a tmp file and use str_replace_editor or `grep -n -B <lines before> -A <lines after> <query> <filename>` to confirm output. * When viewing a page it can be helpful to zoom out so that you can see everything on the page. Either that, or make sure you scroll down to see everything before deciding something isn't available. * When using your computer function calls, they take a while to run and send back to you. Where possible/feasible, try to chain multiple of these calls all into one function calls request. * The current date is {datetime.today().strftime('%A, %B %-d, %Y')}. </SYSTEM_CAPABILITY> <IMPORTANT> * When using Firefox, if a startup wizard appears, IGNORE IT. Do not even click "skip this step". Instead, click on the address bar where it says "Search or enter address", and enter the appropriate search term or URL there. * If the item you are looking at is a pdf, if after taking a single screenshot of the pdf it seems that you want to read the entire document instead of trying to continue to read the pdf from your screenshots + navigation, determine the URL, use curl to download the pdf, install and use pdftotext to convert it to a text file, and then read that text file directly with your StrReplaceEditTool. </IMPORTANT>"""

@jinen @reha congratulations!

NG DB needs NG UI, and notebooks are so much nicer and more useful than the old workbenches. Great job @duckdb and @motherduck !

Introducing analytics for Twitter Cards blog.twitter.com/2014/introd…

Great paper for the technical ones and great summary for everyone, highly recommend!

Nice change from Gemini, will remove a lot of lines of code... Structured output support (like Pydantic) is kind of must-have for any LLM API provider now.

Great point. So basically a search engine in the decision process space. Your mind can only cover so much and LLM may discover some missing paths for you. Some of the paths may fall in the hallucination category since LLM can't really judge, but they usually pretty easy to spot.

I agree most of the statements, except for that everyone will interact with a chat interface for the "just-in-time" code. Human language is a lossy abstraction of the real intention and ten people saying the same words may have ten different intentions, even with personalized histories. That's why product manager is a real job and good product managers are hard to come by. We will have a new tool to get these "just-in-time" programs to do a lot of work for sure, just not the "low-code" platforms. Python and JS are the ultimate "no-code" platform, we just need a good UI/UX to use them.

It reminds me of CORBA in terms of functionality, the og agent framework:-) The impact will be much bigger though.

Lol this is really a nice way to learn the dependency hell.

Now StackOverflow is almost gone, where can we get HUMAN-answers for REAL hard questions?

QueryGPT – Natural Language to SQL Using Generative AI uber.com/blog/query-gpt/ Very interesting work from Uber. To get Text2SQL 100% correct is hard. One interesting point in the work is that you still need domain knowledge to narrow down LLM queries to workspaces (basically different types of workload) so that LLM can get relevant info. Efficient retrieval is not LLM's strong suit and traditional IR and database can do a lot of help here.

George R.R. Martin Wants to Kill You for $20,000 mashable.com/2014/06/05/geor… via @mashable

The results from GPT-5 with Cursor is very impressive. This week is really nuts.

Great post, thanks for sharing!

@jimmyjchen hey it is not a waste if you saved your city somewhere:-)

@chuangl4 ha, the game was crazy:-)

Would love to join, thanks!

This is a great move.

@tao @shanjing @SiweiShen you just fold all the time, nothing to be too proud of other than getting a free dinner:-)

RIP winamp techcrunch.com/2013/11/20/af… via @techcrunch

@nimalan lmao, so true! I have six charging cables for my iphone: office/backpack/car/bedroom/livingroom

Hey Arash, that's great to hear👍!

@THISWILLWORK this is more like #twitterhighfive:-) btw good work hacking from home!

@J_ "His French accent makes his talks attractive."

This is kind of reason why Google might be in trouble. I tried the query "What is Contextual Retrieval for RAG?" on @Google and @perplexity_ai , and the results screenshots are below. Which one would you prefer? This query is only one example, you can easily find much more similar ones.

For any SW engineer that is worrying about being replaced by AI: "your job as an engineer to finding the spec, designing the system, and proving that it works is more crucial than ever." -from @lateinteraction And your safest bets: (link in reply, only 6 minutes, highly recommended).

@nimalan congratulations!

Here is the thinking trace of R1 coming up with the section plans (section title + the prompts to generate the section): <think> Okay, let's tackle this query. The user wants a list of topics for an analytical research report on how agentic and generative AI affect non-tech jobs. They provided a bunch of references, so I need to extract the most relevant topics from that content. First, I'll go through each reference to identify key points. The main themes seem to be job displacement, new roles creation, skill adaptation, ethical concerns, industry-specific impacts, economic effects, and governance. The first reference talks about developing technical and durable skills. That points to a topic on skill adaptation. The second mentions risks like bias and misuse, which ties into ethical risks. Another reference highlights job loss stats, so economic impact is a topic. Agentic AI in specific sectors like healthcare, finance, and legal comes up a few times. Those should be separate industry-specific topics. The emergence of new roles like AI trainers or ethicists is mentioned, so that's another topic. The debate between job displacement vs. creation is present in multiple references, so that's a key topic. Workforce transformation strategies, including HR integration and leadership, are also important. I need to ensure each topic is directly related to non-tech jobs. For example, automation in customer service or administrative roles. Also, governance and regulatory challenges are recurring, so that's a must-include. I should avoid topics that are too technical or not directly related to non-tech job impacts. Each topic needs a clear title and a prompt that guides research. Let me structure them to cover different angles: economic, industry-specific, skills, ethics, governance, future roles, etc. Now, checking the JSON structure. The user wants a list of topics with titles and prompts, in English, without markdown. Make sure the JSON is valid, no trailing commas. Each prompt should ask for detailed analysis, relevant stats, examples, and implications. Let me draft the topics, ensuring they're distinct and cover all aspects from the references. Titles should be concise, prompts should guide comprehensive exploration. For example, one on economic impact would include the stats from the references about job losses and new roles. Another on ethical risks would mention bias, security, legal issues. Double-checking that all key points from the references are covered. Industry impacts include finance, healthcare, legal, HR. Skill adaptation includes both technical and soft skills. Governance includes regulatory approaches and safety testing. Alright, compiling the JSON with 10 topics, each with a title and prompt. Make sure the JSON syntax is correct, using quotes properly, commas between objects, and no extra brackets. That should do it. </think> { "topics": [ { "title": "Job Displacement and Workforce Transformation in Non-Tech Sectors", "prompt": "Analyze how agentic AI automates complex decision-making tasks in non-tech roles (e.g., administrative, customer service) and its implications for workforce restructuring. Include statistics on projected job losses, case studies of industries like finance or logistics, and strategies for mitigating displacement risks." }, { "title": "Emergence of New Roles and Skill Requirements in AI-Driven Economies", "prompt": "Examine the creation of non-tech AI-centric roles (e.g., AI trainers, ethics auditors) and evolving skill demands (data literacy, prompt engineering). Discuss education pathways, corporate training programs, and the role of human-AI collaboration in future job markets." }, { "title": "Ethical Risks and Governance Challenges in Autonomous AI Systems", "prompt": "Evaluate risks such as algorithmic bias, security vulnerabilities, and legal liabilities in agentic AI deployment. Propose governance frameworks for ensuring accountability, transparency, and fairness in AI-augmented workplaces across sectors like healthcare and law." }, { "title": "Economic Impact of AI on Global Labor Markets", "prompt": "Assess macroeconomic effects including wage polarization, regional disparities in job vulnerability, and shifts in labor demand. Incorporate data on projected career changes (2030 timelines), impacts on lower-income workers, and balance between productivity gains vs. inequality." }, { "title": "Industry-Specific Disruption: Case Studies in Finance and Healthcare", "prompt": "Compare how agentic AI transforms compliance, fraud detection, and diagnostics in finance/healthcare. Analyze workforce adaptation strategies, ROI metrics from early adopters, and ethical dilemmas in AI-driven decision-making processes." }, { "title": "Human-AI Collaboration Models in Non-Technical Workflows", "prompt": "Explore frameworks for integrating agentic AI as 'digital assistants' in HR, sales, and operations. Discuss UiPath’s RPA-AI synergy, Visier’s analytics platform, and challenges in maintaining human oversight while optimizing efficiency." }, { "title": "Psychological and Societal Impacts of AI Workforce Integration", "prompt": "Investigate worker anxiety, skill obsolescence fears, and generational disparities in AI adoption. Address freelancers’ adaptive strategies, changes in work-life balance, and societal responses to potential AGI-driven labor obsolescence." }, { "title": "Regulatory and Policy Responses to AI-Driven Job Market Shifts", "prompt": "Critique current AI safety testing protocols, economic safeguards, and proposals like resource reservation for humans. Assess the feasibility of international governance models and corporate accountability in workforce transitions." }, { "title": "Strategic Adaptation for Organizations and Leadership", "prompt": "Outline actionable steps for businesses to pilot agentic AI, redesign workflows, and upskill employees. Reference Deloitte’s 2024 ROI findings, NTT DATA’s implementation strategies, and leadership challenges in ethical AI deployment." }, { "title": "Future Scenarios: From Agentic AI to Artificial General Intelligence (AGI)", "prompt": "Speculate on long-term impacts of AGI surpassing human decision-making capabilities in non-tech domains. Discuss existential risks, redefined human purpose in labor markets, and preparatory measures for radical economic model shifts." } ] }

My bet would be the first one: 1. attention != memory, if you still have to do context engineering, that's already outside of the LLM's realm. I think the current transformer architecture determined it can only work well on local solutions ; 2. taste can't be easily added to probabilistic models producing most likely outcome based on popularity votes. Good architecture design needs a lot of compromises and sometimes bets on undetermined constraints, which can't really be inferred from previous knowledge. And I would add "dynamically" to composable in the first one to make it clearer: Create something composable dynamically -> Software on-demand supported by well-designed infra to provide guardrails, utilities, efficiencies

Former Twitter Chief Scientist Launches Steven, An Emoji-Based Social App techcrunch.com/2014/09/04/fo… via @techcrunch

Tweet cake

so cold

@tianhonghe I should have mentioned Copyright @wangtian

so close...

Twitter details how new home-grown system coordinates data analytics gigaom.com/2014/06/27/twitte… via @gigaom

@THISWILLWORK and spilled ketchup of size 1TB all over the place?

Original post here: blog.pragmaticengineer.com/s…

Workampers -- Welcome to Amazon Town - Yahoo! Finance finance.yahoo.com/news/welco… via @YahooFinance

Damn, 61 of them! I just need one really. Thanks @johnrushx for the great list and summary.

You know you are in trouble when you see 21 lines of comments explaining a boolean variable.

Just added a new function to ask.py: it now can extract structured information from web search results or URLs you specify. (Code here github.com/pengfeng/ask.py) Here is a list of LLM-GenAI startups it extracts from a simple "LLM Gen-AI Startup" web search (only showing the first a few lines because of length limit, there are total 160 items with duplicates from the top 10 search result pages): name,description,SourceURL Cohere,Cohere is an AI startup that builds multilingual LLMs for enterprise businesses to streamline tasks.,explodingtopics.com/blog/gen… Hugging Face,"Hugging Face is a collaborative AI community that creates tools for developers, offering over 61,000 pre-trained models and 7,000 datasets.",explodingtopics.com/blog/gen… Tabnine,Tabnine is an AI assistant for software developers that uses generative AI to predict or suggest the next lines of code.,explodingtopics.com/blog/gen…

And the step-by-step tutorial is here: github.com/leettools-dev/lee…

Great post and should save a lot of time and tokens. Stating that LLMs have limitations does not mean that they can't do great (and magical) things right now. I really believe the most important thing for AI developers right now is to learn the TRUE limitations of the current Transformer-based architecture, so that we do not waste time doing things LLMs can do well and spend time on things LLMs need help to deliver their potentials. From the post: Implications for AI Development This framework reshapes expectations: ❌ What Won’t Work Recursive self-improvement via self-prompting Assuming more tokens = more intelligence Agent loops without grounding ✅ What Will Work Tool-augmented reasoning with access to fresh data Multi-agent verification systems with cross-checking Environment-grounded tasks that introduce external input Hybrid architectures combining neural and symbolic logic Constitutional AI with structured external feedback loops

The PR article is here: theverge.com/2024/11/27/2430… And the web site sdsa.ai/

From visual studio - vi - emacs - eclipse - intellij - vscode, wondering who will be the last one standing this time ...

Prompt 1: How will agentic AI and generative AI affect our non-tech jobs? Source: teddit.net/r/ArtificialIntel… Answers: chatgpt.com/share/67a6a4db-1… g.co/gemini/share/d63f48b939…

I wrote a one-program query engine here github.com/pengfeng/ask.py And the result it generated for the same query here: huggingface.co/spaces/LeetTo… I think it is not too bad either: # Answer Contextual Retrieval for Retrieval-Augmented Generation (RAG) is a method designed to enhance the retrieval accuracy of knowledge systems by addressing the issue of context loss that typically occurs when documents are chunked for embedding. This approach utilizes "Contextual Embeddings," which append chunk-specific contextual information to each segment before it is processed. As a result, the retrieval effectiveness is significantly improved, reducing failures by up to 49% and, when combined with reranking, by 67%[1][2]. This method incorporates two main sub-techniques: Contextual Embeddings and Contextual BM25, allowing for more accurate retrieval of information from knowledge bases, particularly in application areas requiring precise context, such as customer support or legal analysis[4][5]. Therefore, Contextual Retrieval not only helps maintain the necessary context but also improves the reliability of the RAG systems overall[2][6]. # References [1] techcommunity.microsoft.com/… [2] anthropic.com/news/contextua… [3] machinelearning.apple.com/re… [4] anthropic.com/news/contextua… [5] techcommunity.microsoft.com/… [6] anthropic.com/news/contextua… [7] techcommunity.microsoft.com/… [8] m-ruminer.medium.com/anthrop… [9] anthropic.com/news/contextua… [10] news.ycombinator.com/item?id…

Run your own version of Perplexity in one single file - Part 3: Chonkie and Docling The idea is to show how the search-extract-summarize process works in AI search engines such as Perplexity. The code is open sourced here: github.com/pengfeng/ask.py Just got some time to add the newly released Chonkie chunker (github.com/bhavnicksm/chonki…) and Docling document converter (github.com/DS4SD/docling) to the process. So the program can query against local PDFs now: - put your PDF files under the 'data' subdirectory (we have a demo Readme as an example) - run: python -c -i local -q 'how does Ask.py work?'

Write your own version of Perplexity in an hour I wrote a simple Python program (around 250 lines) to implement the search-extract-summarize flow, similar to AI search engines such as Perplexity. Code is here: github.com/pengfeng/ask.py Basically, given a query, the program will - search Google for the top 10 web pages - crawl and scape the pages for their text content - chunk the text content into chunks and save them into a vectordb - performing a vector search with the query and find the top 10 matched chunks - use the top 10 chunks as the context to ask an LLM to generate the answer - output the answer with the references Of course this flow is a very simplified version of the real AI search engines, but it is a good starting point to understand the basic concepts.

...

@akashgarg @twitter congratulations!

Mark Zuckerberg's Chinese interview -- in Mandarin usat.ly/1wlr6RO via @usatoday