[dongecko]

The company I work for has tons of documentation and regulations for several areas. In some areas the documents are well over a thousand and for the ease of use of these documents we build RAG based chat bots. This is why I have been playing with RAG systems on the scale of "build completely from scratch" to "connect the services in Azure". The retrieval part of a RAG is vital for good/reliable answers and if you build it naive, the results are not overwhelming.

You can improve on the retrieved documents in many ways, like - by better chunking,

- better embedding,

- embedding several rephrased versions of the query,

- embedding a hypothetical answer to the prompt,

- hybrid retrieval (vector similarity + keyword/tfidf/bm25 related search),

- massively incorporating meta data,

- introducing additional (or hierarchical) summaries of the documents,

- returning not only the chunks but also adjacent text,

- re-ranking the candidate documents,

- fine tuning the LLM and much, much more.

However, at the end of the day a RAG system usually still has a hard time answering questions that require an overview of your data. Example questions are:

- "What are the key differences between the new and the old version of document X?"

- "Which documents can I ask you questions about?"

- "How do the regulations differ between case A and case B?"

In these cases it is really helpful to incorporate LLMs to decide how to process the prompt. This can be something simple like query-routing, or rephrasing/enhancing the original prompt until something useful comes up. But it can also be agents that come up with sub-queries and a plan on how to combine the partial answers. You can also build a network of agents with different roles (like coordinator/planner, reviewer, retriever, ...) to come up with an answer.

* edited the formatting

[CharlieDigital]

> You can also build a network of agents

My experience has been that they are far too unpredictable to be of use.

In my testing with agent networks, it was a challenge to force it to provide a response, even if it was imperfect. So if there's a "reviewer" in the pool, it seemed to cause the cycle to keep going with no clear way of forcing it to break out.

3.5 actually worked better than 4 because it ran out of context sooner.

I am certain that I could have tuned it to get it to work, but at the end of the day, it felt like it was easier and more deterministic to do a few steps of old-fashioned data processing and then handing the data to the LLM.

[dongecko]

That is an interesting observation. I have not gotten to the point of too long cycles and I can think of two reasons for that.

Maybe my use case is narrow enough, so that in combination with a rather constraining and strict system message an answer is easy to find.

Second, I have lately played a lot with locally running LLMs. Their answers often break the formatting required for the agent to automatically proceed. So maybe I just don't see spiraling into oblivion, because I run into errors early ;)

[CharlieDigital]

The use case we have is that we are asking the LLM to write articles.

As part of this, we tried having a reviewer agent "correct" the writer agent.

For example, in an article about a pasta-based recipe, the writer wrote a line like "grab your spoon and dig in" and then later wrote another line about "twirl your fork".

The reviewer agent is able to pick up this logical deviation and ask the writer to correct it. But given an instruction like "it doesn't have to be perfect", the reviewer will continue to find fault with the output from the writer for each revision so long as the content is long enough.

One workaround is that instead of fixing one long article, have the reviewer only look at small paragraphs or sections. The problem with this is that the final output can feel disjointed since the writer is no longer working with the full context of the article. This can lead to repeated sentence structure or even full on repeated phrases since you're no longer applying the sampling settings across the full text.

In the end, it was more efficient and deterministic to simply write two discrete passes: 1) writer writes the article and 2) another separate call to review and correct.

[TheLegace]

How do you get the output to be formatted correctly or without any branches.

Say for example I want a step-by-step instruction for an action.

But the response will have 1. 2. 3. and sometimes if there are multiple pathways there will long answer with 2.a,b,c,d. This is not ideal I would rather have the most simple case(2.a.) and a short summary for other options. I have described it in the prompt but still cannot get nice clean response without to many variations of the same step.

[dongecko]

I have not encountered this problem yet. When I was talking about the format of the answer I meant the following: No matter if you're using Langchain, Llamaindex, something self made, or Instructor (just to get a json back); under the hood there is somewhere the request to the LLM to reply in a structured way, like "answer in the following json format", or "just say 'a', 'b' or 'c'". ChatGPT tends to obey this rather well, most locally running LLMs don't. They answer like:

> Sure my friend, here is your requested json:

> ```

> {

> name: "Daniel",

> age: 47

> }

> ```

Unfortunately, the introductory sentence breaks directly parsing the answer, which means extra coding steps, or tweaking your prompt.

[int_19h]

It's pretty easy to force a locally running model to always output valid JSON: when it gives you probabilities for the next tokens, discard all tokens that would result in invalid JSON at that point (basically reverse parsing), and then apply the usual techniques to pick the completion only from the remaining tokens. You can even validate against a JSON schema that way, so long as it is simple enough.

There are a bunch of libraries for this already, e.g.: https://github.com/outlines-dev/outlines

[PeterisP]

If that's what you need, it would make all sense to redo the instruction fine-tuning of the model, instead of fiddling with prompt or processing to work around the model settings that go counter to what you want.

[dongecko]

At the very beginning of my journey I did some fine tuning with Lora on a (I believe) Falcon model, but I haven't looked at it since. My impression was that injecting knowledge via fine tuning doesn't work, but tweaking behavior does. So your answer makes much sense to me. Thanks for bringing that up! I will definitively try that out.

[chenxi9649]

Interesting, it seems that using an LLM as an agent to help with knowledge retrieval is one concrete use case that I've seen people do repeatedly.

It also feels like we are at a bottle neck when it comes to the knowledge retrieval problem. I wonder if the "solution" to all of these is just a smarter foundational model, which will come out of 100x more compute, which will cost approximately 7 trillion dollars.

[dongecko]

I also think of the retrieval part as a bottleneck and I am super excited of what the future holds.

In particular, I wonder if RAG systems will soon be a thing of the past, because end to end trained gigantic networks with longer attention spans, compression of knowledge, or hierarchical attention will at some point outperform retrieval. On the other hand, I can also see a completely different direction coming, where we develop architectures that, like operating systems, deal with memory management, scheduling and so on.