Knowledge Graphs in Retrieval Augmented Generation with WhyHow.AI
0 likes209 views
Whyhow.ai enhances retrieval-augmented generation (RAG) processes using knowledge graphs, enabling developers to create and manage domain-specific graphs easily. The approach addresses challenges in understanding context and relevance, allowing for more accurate and comprehensive responses to complex queries. Their solution focuses on quick graph creation and management, ensuring efficient interactions between agents and small, well-scoped graphs.
![Demo
WhyHow
...
],
"patterns": [
{
"head": "character",
"relation": "casts",
"tail": "spell",
"description": "A character casts a specific
spell, e.g., Harry casts Expelliarmus."
},
{
"head": "character",
"relation": "goes_to",
"tail": "location",
"description": "A character goes to a
location, e.g., Hermione goes to Hogwarts."
},
{
"head": "character",
"relation": "uses",
"tail": "magical_object",
"description": "A character uses a magical
object, e.g., Ron uses the Invisibility Cloak."
}
]
}
+ =](https://image.slidesharecdn.com/whyhow-zilliz-meetup-240528212412-f6d892c7/85/Knowledge-Graphs-in-Retrieval-Augmented-Generation-with-WhyHow-AI-13-320.jpg)
Knowledge Graphs in Retrieval Augmented Generation with WhyHow.AI
- 1. WhyHow.AI How WhyHow augments RAG with Knowledge Graphs WhyHow
- 2. About Chris ● Background as a SWE, SA, PM, EM, prev. founder ● Deep focus on platform engineering and developer experiences ● My motivating reason for working in this space: ○ I love developer experiences that are simple, intuitive, and controllable… Chris …but building in Gen AI felt anything but WhyHow
- 3. About WhyHow WhyHow builds developers tooling that makes it fast and easy to create, manage, and orchestrate knowledge graphs in your RAG pipelines to make AI more accurate, deterministic, and explainable. Tom Chris Chia WhyHow
- 4. Vanilla RAG ● Split documents into chunks & pre-process ● Generate embeddings and upsert to vector store ● Use natural language to retrieve similar chunks ● Give chunks to LLM to generate response WhyHow
- 5. Challenges we faced with basic RAG ● Models don’t understand my domain context ○ Does ‘vehicular capacity’ mean # of cars on a road, or # of passengers in a car? ○ What is ‘rice’ to a farmer? A chef? A nutritionist? A grocery store owner? ● Accounting for varied queries ○ “I want to go to an Italian beach..” vs “I’m stressed and need a vacation…” ● Similarity != relevance ○ What is a beach house vs a beachfront house? ● Difficult to build comprehensive responses ○ “Who are all the LPs in my fund?” ● Handling complex queries ○ “Who are all the LPs in my fund that have invested at least $10M and have special data access rights?” WhyHow
- 6. Improving RAG ● Prompt improvement ● Chunking strategies ● Embedding models ● Sub queries ● Hybrid search ● Re-ranking ● Agents ● HYDE ● …. WhyHow
- 7. Our approach Put your data in small, well-scoped graph, and let the LLM talk to it. WhyHow
- 8. Our Insights ● Many small graphs > one massive graph ○ Build a graph scoped only to what the agent needs to solve the problem. ● LLMs help you do this fast ○ So use them to quickly iterate through graph creation, schema generation, etc. ● Graphs should be mapped to your view of the world ○ Graphs should be scoped to your domain and tasks. ● Knowledge graphs + Vector databases = better RAG ○ Graphs and vector databases are force multipliers for each other. WhyHow
- 9. Some benefits of augmenting RAG with small KGs ● Structured grounding of answers ○ Give LLMs only domain-specific, finite context. ● Completeness of answer ○ Difficult to ensure comprehensive answers with top_k alone. ● Complex multi-hop querying ○ Tell me about all the LPs who have invested at least $10M and have special data access rights. WhyHow
- 10. But… It is difficult and time consuming to do comprehensive graph creation and management. WhyHow
- 11. Our solution Make a small graphs really quickly, and iterate until you have a good enough representation of your domain to solve your problem. WhyHow
- 12. Our offerings ● Graph creation ○ We help customers create graphs in 3 different ways right now: 1. Questions 2. Schemas 3. Structured CSVs ○ Chunk linking to link graphs to raw text ● Graph management ○ Schema management, graph management, backups, versioning, access control, and several other things that make a graph usable for an organization ● Orchestration ○ Complex queries, graph to graph tasks, and solutions that make it easy for agents to talk to many small graphs WhyHow
- 13. Demo WhyHow ... ], "patterns": [ { "head": "character", "relation": "casts", "tail": "spell", "description": "A character casts a specific spell, e.g., Harry casts Expelliarmus." }, { "head": "character", "relation": "goes_to", "tail": "location", "description": "A character goes to a location, e.g., Hermione goes to Hogwarts." }, { "head": "character", "relation": "uses", "tail": "magical_object", "description": "A character uses a magical object, e.g., Ron uses the Invisibility Cloak." } ] } + =
- 14. WhyHow x Zilliz: rule-based-retrieval package https://github.com/whyhow-ai/rule-based-retrieval WhyHow
- 15. Recap ● Build graph to represent the full scope of the question, and nothing further ○ Reduce risk of context poisoning ● Represent information according to how domain experts interact with the domain ○ Is this a rice graph for Farmers? Chefs? Nutritionists? Grocery store owner? ● Let agents talk to these graphs to perform small, scoped tasks ○ Small agent microservices talk to small graphs ● KGs + vector databases are force multipliers for each other WhyHow
- 16. Thank you! ● Our Website - https://www.whyhow.ai ● Our Blog - https://medium.com/enterprise-rag ● Discord - discord.gg/twcFcaezc3 WhyHow