Back office ops · Production

How Reforge Built Their AI-Powered Browser Extension Using RAG and Chain of Thought

The problem

Reforge users consistently struggled to apply course content to their actual daily work, even after Reforge launched Artifacts to bridge the gap.

First attempt

The initial simple RAG setup did not produce reliable outputs, and the LLM failed to accurately classify document types, resulting in irrelevant suggestions being shown to beta testers.

Workflow diagram · grounded in source
1
User requests document help
trigger
“When the user clicks "Help Me Improve My Document", the contents of their document are retrieved with JavaScript. A request is made to the Reforge backend with the document contents and the type of help the user wants.”
2
Embedding and similarity search
ai_action
“Embeddings are run on the user's document and their prompt, and a similarity search is performed using these embeddings. Relevant Reforge material is found across artifacts, guides, and lessons.”
3
Document type classification
ai_action
“we decided to explicitly add a step to classify the document. This allowed us to direct the LLM along a specific path”
4
Parallel suggestion generation
ai_action
“We asked the LLM in parallel to produce a suggestion given the more specific Reforge content and the specific questions. We then collected all the suggestions and streamed all three of them back to the extension.”
5
Stream suggestions to extension
output
“streamed all three of them back to the extension”
6
Track recommendation distribution
feedback_loop
“We do store the distribution of course recommendations. When we analyzed the distribution of course recommendations provided by the Reforge extension”
Reported outcome

After upgrading from simple RAG to a Chain of Thought approach with explicit document classification, the distribution of course recommendations diversified and less than 50% of responses contained no Reforge references, with further increases in references seen after upgrading to version 4.0.

Reported metrics
responses with no Reforge referencesless than 50%
Reforge references per response after v4.0 upgradesignificant increase in references
RAG output reliabilitydid not produce reliable outputs consistently
Reported stack
PineconeLaunchDarklyAdalineSegmentMetaBaseAmplitudeSnowflakeRetoolGoogle DocsNotionConfluenceCoda
Source
https://www.reforge.com/blog/howwebuiltit
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After upgrading from simple RAG to a Chain of Thought approach with explicit document classification, the distribution of course recommendations diversified and less than 50% of responses contained no Reforge referenc…

What tools did this team use?

Pinecone, LaunchDarkly, Adaline, Segment, MetaBase, Amplitude, Snowflake, Retool, Google Docs, Notion.

What results were reported?

responses with no Reforge references: less than 50%; Reforge references per response after v4.0 upgrade: significant increase in references; RAG output reliability: did not produce reliable outputs consistently (source-reported, not independently verified).

What failed first in this deployment?

The initial simple RAG setup did not produce reliable outputs, and the LLM failed to accurately classify document types, resulting in irrelevant suggestions being shown to beta testers.

How is this back office ops AI workflow structured?

User requests document help → Embedding and similarity search → Document type classification → Parallel suggestion generation → Stream suggestions to extension → Track recommendation distribution.