@nick_h.near [Posted on DevHub](https://near.social/#/devgovgigs.near/widget/Post?id=1731) ## Solution: Multichannel developer support Assistant ###### Requested amount: 18396 USDT ###### Requested sponsor: @neardevgov.near As the ecosystem around blockchain and decentralized platforms grows, the need for an intuitive, precise, and comprehensive system to access development-related knowledge is more critical than ever. We propose a solution to make NEAR's vast developer documentation seamlessly queryable and discoverable for open-source developers. Why it's matter: if we want to onboard more builders to the space, we need to empower them to build their products instantly without wasting a second on searching docs or waiting on manual support. 1. **Solution components (part of current proposal):** 1. Gathering & rewriting/labeling existing knowledge base, definning optimal sources; 2. Vector embeding and linking it to LLM; 3. End-user integrations (Discord/TG/Webapp integration with BOS portal) with ability to alarm on issues with AI assistance; 4. Collecting analytics and user feedback to improve the quality of AI-powered support assistants; Technical Overview of our Approach 2.1. Core Tech Stack Language Model: GPT-3.5 / GPT-4.0 by OpenAI Vector Embedding & Retrieval: lightweight vector database + pinecone Other data document Storage: MongoDB **2.2. Unique RAG Workflow** Our RAG process is custom-tailored to ensure the highest degree of accuracy, relevance, and coherence in responses. Manual Re-writing of Developer Documents: Unlike generic web scrapers, we manually rewrite and format all developer documentation. This ensures the exclusion of irrelevant data and maximizes the quality of the information source. Enhanced with clear titles, headers, sections, bullet points, and code snippet indicators, this meticulous approach distinguishes our dataset from any other. Optimized Chunking: Document contents are divided into chunks of 1500 characters, ensuring comprehensive context without overwhelming the model. Semantic Vector Embedding: Each chunk is embedded into a vector and stored in either Pinecone or custom-made vector DB's database, leveraging cosine similarity to retrieve the most relevant document fragments. Dynamic Query Processing: When a developer poses a query, it's vector-embedded and compared with our database, fetching the top 5 most relevant chunks. Tailored Prompt Design for LLM: Extracted chunks are structured into a prompt, guiding GPT-3.5 or 4 to deliver concise, accurate, and relevant answers while leveraging code snippets or further reading links when appropriate. **3. Distinct Advantages Over Other RAG Providers** Human-Curated Data Source: Our manual rewriting and formatting eliminates noise, resulting in a high-quality primary data source significantly improving response accuracy. Contextual Formatting: By adding clear demarcations like titles, headers, sections, and bullet points, we enable the model to better understand and differentiate between different types of content. This facilitates more nuanced and relevant answers. Advanced Prompt Design: Our prompts are meticulously crafted, guiding the LLM not just in content but in style, tone, and depth, ensuring that answers are not only accurate but also developer-friendly. Iterative Parameter Tuning: Leveraging feedback loops and iterative testing, we constantly refine our system parameters (chunk size, retrieval count, model temperature) to optimize the quality of generated responses further. **4. Upcoming tech implementation:** 1. Feedback Integration: Implement a mechanism to allow developers to provide feedback on the generated answers, allowing us to refine our process based on real-world user experience continually. 2. Adaptive Learning: Incorporate a system where frequently queried topics inform periodic retraining or fine-tuning of the model, ensuring that the most sought-after information is always at the forefront. 3. Expanding Contextual Elements: Introduce meta-tags and categories for each chunk to provide an additional layer of context, which can be used to further refine search results and answer generation. 4. Query Suggestion System: If a developer's query does not yield satisfactory results, our system can suggest alternative queries or topics they might explore, enhancing discoverability. **5. Use cases we target to assist:** - Dev Rel and support automation - By automating essential support, we aim to improve developers' experience on NEAR; - Providing cross-channel support to support community with tech-related issues whenever they need it; - Support ecosystem hackathons; - Providing data on developer requests/issues that will help to prioritize Improvements of tech documentation and/or features that are missed; **5. Idea/Vision Validation** Our team is an alumnus of the latest Outlier Ventures acceleration program (July 2023), and recently, we joined the first cohort of the NEAR Horizon program designed to support founders on their journey from product building to doing sustainable business. We are excited with the quality of the program and the feedback we got there, as well as the effort Laura Cunningham and Jarrod Barnes put into making it truly valuable for the NEAR ecosystem and every team in the cohort; I have discussed our approach with several people from the NEAR ecosystem: 1. Illia Poloskuhin (Founder of NEAR) over the AI-focused event in NYC on the 20th of Sep; Output - to start with this proposal, 2. Boris Polania (Dev Rel at DevHub) - over the ETHGlobal in NYC Outputs: refs for high-quality proposals & validation of the problem with manual support; 3. Cameron Dennis (CEO of Banyan) - discussed the potential integration of a bot to the BOS webpage to support the developer community; Output: refs to Paul from Mintbase to discuss tech implementation; 4. Paul Kuveke (COO at Mintbase) - discussed the solution and its implementation and created an additional TG chat with Micro - the head of innovation who leads AI R&D at Minbase. Output - discussed our differentiation and tech implementation. Micro liked our uniques approach, additional communication regarding further integrations with Mintbase in progress; 5. Maggie Sun (BD team at NEAR Foundation) - discussed the product vision and our G2M strategy and how NEAR Foundation may additionally support us in the long term; Output: agreed to have an additional follow-up call to define actionable steps; 6. Demo call with Jarrod Barnes (Head of ecosystem at NEAR Foundation) and Ken Miyachi (Technical Lead at NEAR Foundation); We got positive feedback on our approach, and agreed to push this proposal; **6. Path Towards Sustainability** In the case this proposal gets approved, we would deploy all product elements and work with the NEAR ecosystem as a design partner that will help us to refine the product and make it commercially ready; Our stack might be used later for supporting Devs across many chains and unlocking seamless switching between ecosystems. It might be valuable for the ETH tech community to start building on NEAR as well as for the potential NEAR tech community to contribute to ETH projects by using AI assistance. **7. The budget and financing breakdown are as follows:** Server Hosting, Database, High Availability, Backups - $1,000 / 2m SaaS - $398 / 2m: Google Suite - Communication ($80/mo) Calendly - Communication ($49/mo) Adobe Suite - Tools ($30/mo) GitHub Co-pilot - Dev Tools ($40/mo) Team - $17,000 / 2m Total: $18,398 **8. Team:** The team consists of 6 persons based in the UK, Ukraine, Spain, and Canada and have been working together since 2021 The salaries below include two months of platform development and NEAR onboarding, local and corporate taxes. Assisterr does not rent an office; Each team member is working on Assisterr full-time: Nick Havrylial - co-founder, Product: $0,0 Calson Sheng - co-founder, Tech lead: $5,000/2months Dima Dimenko - UX/UI Designer: $3000/2months Tim Kotov - ChatBot Dev: $2,000/2months Den - Front-end $4,000/2months Nazar - Data labeler: $3,000/2months Sum: $17,000 **9. Timeline:** Milestone 1 (1m): – Data sources mapping, data labeling, and vectorisation; Milestone 2 (2w): – End user integrations, testing, collecting feedback; Milestone 3 (4w) – Refining the quality and ecosystem integrations; **10. User testing:** We are going to rely on initial feedback from teams who are a part of NEAR Horizon cohort #1 and ready to use the bot while building their solutions; We expect to provide access to the AI assistant to the community participating in the NEAR hackathon before the NEARCON 2023; The next step will be making the solution viral among the community by creating successful stories in collaboration with Developers who are using our bots to accelerate their dev process; In later stages, after user tests and positive feedback, we also expect to get support from the Foundation by promoting Assisterr multi-channel Assistants at the community resources and including them in the BOS portal; Based on the results and gathered feedback from the community and DevRel team, we will prepare an additional proposal that includes improvements, maintenance, DevRel team-specific requests, and ideas on deeper integration with NEAR ecosystem projects.; **11. Reporting Structure & Payment Schedule** To ensure the Assisterr team meets the expectations of DevHub, we propose to break up the proposed funding into two payments: Milestone 1 ($9,199): Oct 2023 Milestone 2 ($9,199): Nov 2023