Unstruk Data (@unstruk) header
Unstruk Data (@unstruk) profile picture

Now launched, our new Graphlit Platform. API-first developer platform for building knowledge-driven AI apps. @graphlit

Seattle, WA
Joined December 2020
Photo by Unstruk Data (@unstruk)Photo by Unstruk Data (@unstruk)Photo by Unstruk Data (@unstruk)Photo by Unstruk Data (@unstruk)Photo by Unstruk Data (@unstruk)
🎉 It's launch day for @graphlit! 🎉 youtu.be/5C_ltw-Ja_8 Graphlit is an API-first developer platform for building applications with LLMs. If you struggle with #langchain or #llamaindex in production, try us: graphlit.com Simple to integrate, scalable, secure.
2
6
6
1,439
Unstruk Data retweeted
In the latest release of @graphlit, Slack audio and text alerts are now supported for AI app devs. See our changelog: changelog.graphlit.dev/febru… And our API documentation: docs.graphlit.dev/api-refere…
2
3
475
Unstruk Data retweeted
Multi-modal models are 🔥 Simply by using GPT-4 Vision to analyze images, you can publish a written inspection report with LLMs. Read about how developers can use @graphlit to build AI apps for visual inspections. graphlit.com/blog/multimodal…
2
3
363
Unstruk Data retweeted
LLMs can be used to unlock the knowledge from podcasts. We have a new deep dive blog post on using @graphlit for podcast summarization and conversational chatbots for listeners. graphlit.com/blog/llms-for-p…
2
6
1,252
Unstruk Data retweeted
With this crazy OpenAI weekend, several companies have been talking about model routers/gateways for switching apps to different providers. TL;DR: It’s not that simple. Switching between Azure hosted OpenAI and native @OpenAI APIs is easy, given the APIs and models are the same. Swapping to Anthropic Claude works well, but given the preference for XML tags for guiding instructions, it can take slightly more prompt engineering. Pro tip: using XML for guidance is compatible with Anthropic and OpenAI models. The problem we have seen @graphlit is trying to swap GPT models with OSS models like Llama 2. Given our JSON guardrails approach for reliable output from LLMs, Llama 2 won’t “listen” to guidance to return JSON reliably. Differing prompt semantics makes swapping between OpenAI and OSS models non-trivial. As more OSS models become fine-tuned for JSON output, the gap may shrink, but for now OpenAI and Anthropic models are the best at following instructions - and have more potential for model fallback/routing.
5
3
1
11
1,624
Unstruk Data retweeted
We are going to build something new & it will be incredible. Initial leadership (more soon): @merettm @sidorszymon @aleks_madry @sama @gdb The mission continues.
Replying to @sama
I’m super excited to have you join as CEO of this new group, Sam, setting a new pace for innovation. We’ve learned a lot over the years about how to give founders and innovators space to build independent identities and cultures within Microsoft, including GitHub, Mojang Studios, and LinkedIn, and I’m looking forward to having you do the same.
683
2,088
532
21,134
4,660,057
Unstruk Data retweeted
@KirkMarple, Founder and CEO at @unstruk, will dive into Multimodal RAG by using @graphlit and GPT-4 Vision to enable insurance insights. teapawt.org/KirkMarple/status/1722…
Really impressed with the new #OpenAI GPT-4 128K and GPT-4 Vision models. Ingested a dozen images I took of a building fire in Chicago, and generated descriptions, labels, and extracted text into @graphlit See the image of how GPT-4 128K answered my question about the fire.
1
1
1
2
891
Unstruk Data retweeted
Really impressed with the new #OpenAI GPT-4 128K and GPT-4 Vision models. Ingested a dozen images I took of a building fire in Chicago, and generated descriptions, labels, and extracted text into @graphlit See the image of how GPT-4 128K answered my question about the fire.
2
2
2
3
1,087
Unstruk Data retweeted
Interesting comparison between #OpenAI GPT-4V 'low' and 'high' detail modes with @graphlit Low detail (blue background) is still pretty useful, but high appears to give more context. (Also, notice how we're generating labels and description at same time via prompt.) Credit: using diagram from a16z.com/emerging-architectu…
2
4
464
Unstruk Data retweeted
We are excited to announce that we're part of a select group of companies that were accepted into the @deepgram startup program! Graphlit uses Deepgram Nova 2 as our default audio transcription model, which offers excellent performance and quality at low cost.
2
4
369
Unstruk Data retweeted
Simple. Powerful. Graphlit. Build LLM-driven applications, and let us take care of the data ingestion, storage, embeddings, search and LLM conversations via our serverless infra.
KG+LLMs: content for conversations In the next blog in the series, we explain how to use @graphlit and #OpenAI LLMs to build a conversational #knowledgegraph - with just a few GraphQL API calls. Includes faceted queries on extracted entities, and metadata filtering to prompt a relevant 'slice' of the KG. graphlit.com/blog/exploring-…
There's a new version of this post
1
4
501
RT @KirkMarple: KG+LLMs: content for conversations In the next blog in the series, we explain how to use @graphlit and #OpenAI LLMs to bu…
Welcome new agent-powered products in "THE LIST": 🤖 @agentscale 🤖 OpenAgents 🤖 @duckie_ai 🤖 Adala 🤖 @XAgentTeam 🤖 Taxy AI 🤖 @graphlit Check out the complete repository: github.com/e2b-dev/awesome-a…
5
26
1
152
23,416
Unstruk Data retweeted
Extracting useful structured data from unstructured sources, like Reddit, can be difficult. Graphlit makes it simple. We've written a tutorial that walks through the end-to-end API-driven flow for extracting market research data from Reddit posts. graphlit.com/blog/market-int…
1
1
5
113
Extracting useful structured data from unstructured sources, like Reddit, can be difficult. Graphlit makes it simple. We've written a tutorial that walks through the end-to-end API-driven flow for extracting market research data from Reddit posts. graphlit.com/blog/market-int…
1
1
5
113
I wrote an update on the biggest topics of building LLM-powered agents and apps. Topics include security, privacy, memory management, frameworks, plus developers' view. Hope you enjoy the reading, happy to discuss! 🫡 🔗 Link in comment
7
33
3
148
26,899
Unstruk Data retweeted
Yikes @AnthropicAI, somehow I created an infinite loop of responses. I think it was due to wrapping the original user prompt in <user></user> XML tags.
1
1
1
6
281
Unstruk Data retweeted
Tried this out with @graphlit, and used Azure Text Analytics to OCR the workflow diagram, and then asked GPT-3.5 16K some questions about it. Out of the box, it's pretty good. Looking forward to multimodal LLMs, where we can do this in one-shot.
I feel like I intuitiveLy accepted LLMs and why they work, but GPT4Vs ability to understand charts is throwing me off. Like no way you can throw a whole bunch of charts and descriptions into a model and have it learn to understand them, right? There’s a gotta be a layer of orchestration that can leverage segmentation, OCR, etc separately, unless the orchestration is another layer of the model? Does it need to know if it’s a chart or an image? How about an illusion diffusion chart? Bet that would throw it off. Or maybe not. I don’t understand this stuff enough.
1
2
5
631