Open sonaalKant opened 2 months ago
1) Yes, let's pick a tool first. Personally, my favorite is Lucidchart. You can use the free version and then we can get a license if we like it 2) Let's do quick iterations, so you can send us a quick draft so we can converge more quickly
@sonaalKant , @gpsaggese This is the link for the the architecture : https://lucid.app/lucidchart/3e256df2-c19f-4b9b-8478-c13b6f27df83/edit?invitationId=inv_8acc0a33-1c44-4001-bed7-0f6305d98d28
And for the second chart, Is this design good?
They are going in the right direction @Vedanshu7
@sonaalKant, how can we provide better specs? Should we draw on a piece of paper and then take a picture?
The specs for the first one were more like: "read our paper and do some plots explaining the flow, using the style in the fig", instead of converting the pic to LucidChart
The second one is an interesting direction.
@Vedanshu7 / @tkpratardan based on the draft of the paper, anything else you think it can help explain the flow graphically?
FYI @samarth9008
@gpsaggese I have tried to make the flow:
let me know if i am wrong or in which part i need to refine.
Interesting. Let's try to use tools that allow us to create these plots automatically, e.g., https://mermaid.js.org/ or PlantUML We can get someone on fiverr to make the plots look better, if really needed.
Using draw.io is the last resort when we can't use those tools.
The flow of the architecture
LLM vs KG
Drew in Mermaid
Good stuff. Can you also post the source code?
``` mermaid
---
config:
theme: dark
themeVariables:
primaryColor: '#FFF'
secondaryColor: '#FFF'
tertiaryColor: '#000'
primaryFontFamily: Arial
fontSize: 20
---
flowchart LR
sources["Knowledge Sources<br>(Web, Databases,<br>Experts, etc.)"] -- Data --> kgc["Knowledge Graph<br>Construction"]
query["User Query"] -- Query --> qpkge["Query Parsing &<br>KG Extraction"]
dskg["Domain-Specific<br>Knowledge Graph"] -- Data --> qpkge
kgc -- Data --> dskg
qpkge -- Extracted KG --> kgsub["Relevant KG<br>Subset"]
kgsub -- Data --> bnc["Bayesian Network<br>Conversion"]
bnc -- Model --> rrd["Retrieve<br>Relevant Data"]
rrd -- Data --> eval["Evaluate BN Model<br>(e.g., KaizenFlow)"] & preproc["Data Preprocessing<br>& Feature Engineering"]
eval -- Response --> resp["Query Response<br>(with uncertainty<br>estimates)"]
ext_data["External Data Sources<br>(APIs, Streams, etc.)"] -- Data --> eval
preproc -- Model --> training["Model Training &<br>Hyperparameter Tuning"]
training -- Model --> eval
classDef preproc fill:#DCDCDC,stroke:#808080
classDef training fill:#DCDCDC,stroke:#808080
classDef ext_data fill:#DCDCDC,stroke:#808080
style sources stroke-width:2px,stroke-dasharray: 0
style preproc stroke-width:1px,stroke-dasharray: 0
``` mermaid
%%{init: {'theme': 'dark', 'themeVariables': { 'fontFamily': 'Arial, sans-serif', 'fontSize': '16px'}}}%%
graph LR
%% Large Language Models Section
subgraph LLM [Large Language Models]
direction TB
LLM_A[Large Language Models]
LLM_Adv[Advantages_LLM:advantage]
LLM_Lim[Limitations_LLM:limitation]
LLM_A -->|General knowledge| LLM_Adv
LLM_A -->|Language processing| LLM_Adv
LLM_A -->|Generalizability| LLM_Adv
LLM_A -->|Black-box: lack of interpretability| LLM_Lim
LLM_A -->|Implicit knowledge| LLM_Lim
LLM_A -->|Hallucination| LLM_Lim
LLM_A -->|Indecisiveness| LLM_Lim
LLM_A -->|Lack of domain-specific knowledge| LLM_Lim
LLM_A -->|Lack of new knowledge| LLM_Lim
end
%% Knowledge Graphs Section
subgraph KG [Knowledge Graphs]
direction TB
KG_A[Knowledge Graphs]
KG_Adv[Advantages_KG:advantage]
KG_Lim[Limitations_KG:limitation]
KG_A -->|Accuracy| KG_Adv
KG_A -->|Decisiveness| KG_Adv
KG_A -->|Interpretability| KG_Adv
KG_A -->|Domain-specific knowledge| KG_Adv
KG_A -->|Evolving knowledge| KG_Adv
KG_A -->|Incomplete knowledge| KG_Lim
KG_A -->|Lack language understanding| KG_Lim
KG_A -->|Unseen facts| KG_Lim
end
Read paper https://github.com/sorrentum/sorrentum/blob/b2038a09192848d50da5b015dd1ecb6705883d18/papers/fin_econ_ai/fin_econ_ai.pdf
Create a nice figure for the architecture. Sharing with you the current work which we want to improve.
Create a nice figure for KG vs LLM
Resources available:
FYI @gpsaggese