PageRank is one of several interesting centrality metrics that could be applied to a graph to influence RAG on structural data, another one is Triangle Centrality which counts triangles around nodes to figure out their centrality based on the concept that triangles close relationships into a strong bond, where open bonds dilute centrality by drawing weight away from the center:
The paper shows high efficiency compared to other centralities like PageRank, however in some research using the GraphBLAS I and my coauthors found that TC was slower on a variety of sparse graphs than our sparse formulation of PR for graphs up to 1.8 billion edges, but that TC appears to scale better as graphs get larger and is likely more efficient in the trillion edge realm.
This is super interesting! Thanks for sharing. Here we are talking of graphs in the milions nodes/edges, so efficiency is not that big of a deal, since anyway things are gonna be parsed by a LLM to craft an asnwer which will always be the bottleneck. Indeed PageRank is the first step, but we would be happy to test more accurate alternatives. Importantly, we are using personalized pagerank here, meaning we give specific intial weights to a set (potentially quite large) of nodes, would TC support that (as well as giving weight to edges, since we are also looking into that)?
> Here we are talking of graphs in the milions nodes/edges,
That ought to be enough for anybody.
> would TC support that
TC is a purely structural algorithm, it counts triangles so it doesn't take any weights into consideration, but it does return a vector of normalized ranking from 0.0 to 1.0, which you could combine with an existing biasing strategy to boost results that have strong centrality.
Hah indeed, we are doing billion-scale real-time graph rag in louie.ai for fairly regular tasks, so your sentiment resonates ;-)
For something like uploading a big folder of documents, agree with the OP, pretty straightforward, naive in-memory with out-of-the-box embeddings, LLMs, retrieval, and untuned DBs goes far. I expect most vector-supporting dbaas and LLMaaS to be offering in the new year. OpenAI, Claude, and friends are already going in this direction, leaving the rag techniques opaque for now.
(Something folks may not appreciate, and I think is important about what's being done here, is the incremental update aspect.)
1. Lexical search with BM25 alone gives you very relevant results if you can do some work during ingestion time with an LLM.
2. Embeddings work well only when the size of the query is roughly on the same order of what you're actually storing in the embedding store.
3. Hypothetical answer generation from a query using an LLM, and then using that hypothetical answer to query for embeddings works really well.
So combining all 3 learnings, we landed on a knowledge decomposition and extraction step very similar to yours. But we stick a metaprompter to essentially auto-generate the domain / entity types.
LLMs are naively bad at identifying the correct level of granularity for the decomposed knowledge. One trick we found is to ask the LLM to output a mermaid.js mindmap to hierarchically break down the input into a tree. At the end of that output, ask the LLM to state which level is the appropriate root for a knowledge node.
Then the node is used to generate questions that could be answered from the knowledge contained in this node. We then index the text of these questions and also embed them.
You can directly match the user's query from these questions using purely BM25 and get good outputs. But a hybrid approach works even better, though not by that much.
Not using LLMs are query time also means we can hierarchically walk down the root into deeper and deeper nodes, using the embedding similiarity as a cost function for the traversal.
> LLMs are naively bad at identifying the correct level of granularity for the decomposed knowledge. One trick we found is to ask the LLM to output a mermaid.js mindmap to hierarchically break down the input into a tree. At the end of that output, ask the LLM to state which level is the appropriate root for a knowledge node.
> Then the node is used to generate questions that could be answered from the knowledge contained in this node. We then index the text of these questions and also embed them.
Thanks for sharing this! It sounds very interesting. We experimented with a similar tree setup some time ago and it was giving good results. We eventually decided to move towards graphs as a general case of trees. I think the notion of using embeddings similarity for "walking" the graph is key, and we're actively integrating it in FastGraphRAG too by weighting the edges by the query. It's very nice to see so many solutions landing on similar designs!
Something that I'm working on is making it easy to fix spelling and grammatical errors in documents that can affect BM25 and embeddings. So in addition to generating keyword/metadata with LLM, you could also ask it to clean the document; however, based on what I've learned so far, fixing spelling and grammatical errors should involve humans in the process, so you really can't automate this.
Fascinating. I think the process could be automated, though I don't know if it's been invented yet. You would want to use the existing autocomplete tech (probabilistic models based on Levenshtein distance and letter proximity on keyboard?) in combination with actually understanding the context of the article and using that to select the right correction. Actually, it sounds fairly trivial to slap those two together, and the 2nd half sounds like something a humble BERT could handle? (I've heard people getting great results with BERTs in current year, though they usually fine-tune them on their particular domain.)
I actually think even BERT could be overkill here -- I have a half-baked prototype of a keyword expansion system that should do the trick here. The idea is is to construct a data structure of keywords ahead of time (e.g. by data-mining some portion of Common Crawl), where each keyword has "neighbors" -- words that often appear together and (sometimes, but not always) signal relatedness. I didn't take the concept very far yet, but I give it better than even odds! (Especially if the resulting data structure is pruned by a half-decent LLM -- my initial attempts resulted in a lot of questionable "neighbors" -- though I had a fairly small dataset so it's likely I was largely looking at noise.)
It can definitely be automated in my opinion, if you go with a supermajority workflow. Something that I've noticed with LLMs is it's very unlikely for all high-quality LLM models to be wrong at the same time. So if you go by a supermajority, the changes are almost certainly valid.
Having said all of that, I still believe we are not addressing the root cause of bad searches which is "garbage in, garbage out". I strongly believe the true calling for LLM will be to help us curate and manage data, at scale.
> fixing spelling and grammatical errors should involve humans in the process, so you really can't automate this
This is an interesting observation to me. I would have expected that, since LLMs evolved from autocomplete/autocorrect algorithms, correcting spelling mistakes would be one of their strong suits. Do you have examples of cases where they fail?
If you look at my post history, you can see an example of how claude and openai can not tell that GitHub is spelled correctly. The end result won't make a difference but it raises questions regarding how else it can misinterpret things.
At this moment I would not trust AI to automatically make changes.
> 3. Hypothetical answer generation from a query using an LLM, and then using that hypothetical answer to query for embeddings works really well.
I've been wondering about that and am glad to hear it's working in the wild.
I'm now wondering if using a fine-tuned LLM (on the corpus) to gen the hypothetical answers and then use those for the rag flow would work even better.
The technique of generating hypothetical answers (or documents) from the query was first described in the "HyDE (Hypothetical Document Expansion) paper". [1]
Interestingly, going both ways: generate hypothetical answers for the query, and also generate hypothetical questions for the text chunk at ingestion both increase RAG performance in my experience.
Though LLM-based query-processing is not always suitable for chat applications if inference time is a concer (like near-real time customer support RAG), so ingestion-time hypothetical answer generation is more apt there.
We do this as well with a lot of success. It’s cool to see others kinda independently coalescing around this solution.
What we find really effective is at content ingestion time, we prepend “decorator text” to the document or chunk. This incorporates various metadata about the document (title, author(s), publication date, etc).
Then at query time, we generate a contextual hypothetical document that matches the format of the decorator text.
We add hybrid search (BM25 and rerank) to that, also add filters (documents published between these dates, by this author, this type of content, etc). We have an LLM parameterize those filters and use them as part of our retrieval step.
but what about the chunk size, if we have a small chunks like 1 sentence and the hyde embeddings are most of the time larger, the results are not so good
Very interesting. Thank you getting into the details.
Do you chunk the text that goes into the BM25 index?
For the hypothetical answer, do you also prompt for "chunk size" responses?
Very cool and relatable I faced a similar issue for my content categorization engine for local events: http://drophere.co/presence/where (code: https://github.com/itissid/drop_webdemo). Finding the right category for a local event is difficult, an event could be "Outdoorsy" but also "Family Fun" and "Urban Exploration".
My next iteration to solve this problem – I never got to it – was gonna be to generate the most appropriate categories based on user's personal interest, weather, time of day and non PII data and fine-tune a retrieval and a ranking engine to generate categories for each content piece personalized to them.
I organize community calls for Memgraph community and recently a community member presented how he uses hypothetical answer generation as a crucial component to enhancing the effectiveness and reliability of the system, allowing for more accurate and contextually appropriate responses to user queries. Here's more about it: https://memgraph.com/blog/precina-health-memgraph-graphrag-t...
>3. Hypothetical answer generation from a query using an LLM, and then using that hypothetical answer to query for embeddings works really well.
What sort of performance are you getting in production with this one? The other two are basically solved for performance and RAG in general if it is related to a known and pre-processed corpus but I am having trouble thinking of how you don't get a hit with #3.
It's slow. So we use hypothetical mostly for async experiences.
For live experiences like chat, we solved it with UX. As soon as you start typing the words of a question into the chat box, it does the FTS search and retrieves a set of documents that have word-matches, scored just using ES heuristics (eg: counting matching words etc)
These are presented as cards that expand when clicked. The user can see it's doing something.
While that's happening, also issue a full hyde flow in the background with a placeholder loading shimmer that loads in the full answer.
So there is some dead-time of about 10 seconds or so while it generates the hypothetical answers. After that, a short ~1 sec interval to load up the knowledge nodes, and then it starts streaming the answer.
This approach tested well with UXR participants and maintains acceptable accuracy.
A lot of the times, when looking for specific facts from a knowledge base, just the card UX gets an answer immediately. Eg: "What's the email for product support?"
PageRank for better centrality seems neat, but it still doesn't address the probably unsolvable flaw with RAG, the reason why RAG basically can't work.
All RAG DBs under-perform expectations because RAG fundamentally can't find relationships between words necessary to find the information the user cares about. Weird right, isn't this what the 'attention' mechanism is supposed to be good for? It just isn't good enough.
Example: Say you're searching an article and you want to know what occupation a mentioned person has, let's say the person 'Sharon,' is mentioned to have attended several physical chemistry conferences but her occupation is never explicitly mentioned. There's a very good chance every single rag approach will fail to return correct results, will fail to make this connection between 'occupation' attends conference, type of conference and infers 'chemist'. I could go on, but this sort of error is pervasive along all types of information when trying to retrieve with RAG. In the end, solutions like the above seem to just sort of reinvent other query methods, SQL, pagerank etc, with extra steps... there's little point in vectorization at that point...
Isn't this inference an LLM's job? The RAG component just needs to find the Sharon article among a large dataset and pass it (entirely) to the LLM as context.
On the contrary, examples like yours are the entire point of approaches like this one. If you read the HippoRAG paper cited by OP, their motivating example is almost identical to yours, and their evaluations are largely on multi-hop question answering of this kind.
I don't see how this is not possible using knowledge graphs? You retrieve the entity, Sharon, and the additional context you get will be the nodes and edges close to Sharon. After this it becomes the LLM's job because if it is not mentioned in the given context, it should let the prompter know "In the given context the occupation of Sharon could not be found".
This is very cool, I signed up and uploaded a few docs (PDFs) to the dashboard
Our Use case: We have been looking at farming out this work (analyzing complaince documents (manufacturing paperwork) for our AI Startup however we need to understand the potential scale this can operate under and the cost model for it to be useful to us
We will have about 300K PDF documents per client and expect about a 10% change in that document set, month to month -any GraphRag system has to handle documents at scale - we can use S3 as an igestion mechanism but have to understand the cost and processing time needed for the system to be ready to use duiring:
1. inital loading
2. regular updates -how do we delete data from system for example
Super interesting, thanks for sharing. How large a corpus of domain specific text do you need to obtain a useful knowledge graph?
Aider has been doing PageRank on the call graph of code repos since forever. All non trivial code has lots of graph structure to support PageRank. So it works really well to find the most relevant context in the project related to the currently active task.
We have tried from small novels to full documentations of some milion tokens and both seem to create interesting graphs, it would be great to hear some feedback as more people start using it :)
I enjoy Aider, but it has never successfully created a repo map, regardless of whether the codebase is Python, JS, or TS. Are there any plans to allow force-creation and inspection of a repo map?
Thanks, I have tried this, but it simply shows no elements in the repo map with no accompanying errors. Inspecting the file also shows an empty repo map. I'm wondering if Aider needs additional logging when invoking tree-sitter.
I’m on a slightly modified version of 0.52.1 which is getting a bit dated but it works well for me even with not officially supported source, like svelte.
In case this thread helps someone else, some errors with —show-repo-map can be solved by setting environment variable PYTHONIOENCODING=utf-8
This is cool! How is the graph stored and queried? I’m familiar with graph databases, but I don’t see that as a dependency.
Have you tried the sciphi triplex model for extraction? I’ve tried to do some extraction before, but got inconsistent results if I extracted the chunks multiple times consecutively.
The graph is currently stored using python-igraph. The codebase is designed such that it is easy to integrate any graphdb by writing a light wrapper around it (we will provide support to stuff like neo4j in the near future). We haven't tried triplex since we saw that gpt4o-mini is fast and precise enough for now (and we use it not only for extraction of entities and relationships, but also to get descriptions and resolve conflicts), but for sure with fine tuning results should improve.
The graph is queried by finding an initial set of nodes that are relevant to a given query and then running personalized pageranking from those nodes to find other relevant passages. Currently, we select the inital nodes with semantic search both on the whole query and entities extracted from it, but we are planning for other exciting additions to this method :)
Cool idea. IMHO traditional information retrieval is the way to go with RAG. Vector search is nice but also slow and expensive and people seem to use it as magic pixie dust. It works nice for unstructured data but not necessarily that well for structured data.
And unless tuned very well, vector search is not actually a whole lot better than a good old well tuned query. Putting everything together, the practice of turning structured data into unstructured data just so you can do vector search or prompt engineering on it, which I've seen teams do, feels a bit backwards. It kind of works but there are probably smarter ways to get the same results. Graph RAG is essentially about making use of structure of data. Whether that's through SQL joins or by querying some graph database doesn't really matter much.
There is probably some value into teaching LLMs how to query as well; or letting them interface with existing search/query APIs. And you can compensate for poor ranking with larger context sizes and simply fetch a few hundred or even more results with multiple queries. It's going to be a lot faster and cheaper than vector search to scale that.
Looks great. But being burned by other "abstractions", e.g. LangChain, I'm weary of the oversimplification. How are you not going to make those same mistakes?
Do you have any retrieval and generation metric scores (eg, KILT or NQ datasets)?
I know benchmark datasets are not the be-all-end-all, but a halfway decent score and inference-time, would really help sell your framework (or help engineers make the choice).
In any case, very cool work, I built a lot of RAG pipelines as freelance NLP engineer and I will try this out.
Are there any serious LLM engineering communities that have emerged post-hype cycle? Spaces where people are actually pushing boundaries with exploratory engineering, not just theorizing. Somewhere the focus is on testing limits and validating novel approaches - figuring out what's genuinely achievable with this tech.
I assume there's got to be, but I don't have the capacity these days to root around and find it, and I'm genuinely worried about missing out on some really cool shit.
After a lot of experimentation, the only thing that worked in a chat style application is to pass maybe the last 4-5 messages (ideally the entire conversation history) and ask an LLM to summarize the question in the context of the conversation.
Without that it often failed when users asked something like ("Can you expand point 2? , Give a detailed example of the above").
Current implementation(I have 3 indexes) is to provide Query + Past messages and ask an LLM to break it down into
Overall ask:
BM25 optimized question:
Keywords:
Semantic optimized question:
Perform RAG + Rerank and pass the top N passages after this along with the Overall ask in the second LLM call.
If the user asks such a question, your agent should not invoke the RAG at all, but simply answer from the history. You need to focus on your orchestration step.
Search for ReAct agents, can build using either LangGraph or Bedrock Agents.
I might be the wrong target audience (I do have a great interest in this, but I am not doing it at a professional level) but I feel the GitHub could explain things a bit better — now I need to go read someone else's research paper to see what you guys are doing!
(Also readme says see examples folder but it's basically empty?)
Very cool. Have you considered whether incorporating any of the new-ish unsupervised or semi-supervised keyphrase extraction algorithms could give this a boost? Teket (graph-based) and sifrank come to mind, but I’m also wondering if AutoPhrase + an LLM could be powerful.
So I went ahead and tried running the example script with "A CHRISTMAS CAROL" using the "meta-llama-3.1-8b-instruct" and "text-embedding-nomic-embed-text-v1.5" models locally. How long should it take to extract the subgraphs with this kind of setup?
These are knobs that you can tune to make the graph construction more/less opinionated. Generally speaking, the more we make it opinionated the better it fits the task.
At a high-level:
(1) Domain: allows you to "talk to the graph constructor". If you care particularly about one aspect of your data, this is the place to say it. For reference, take a look at some of the example prompts on our website (https://circlemind.co/)
(2) Example Queries: if you know what class of questions users will ask, it'd be useful to give the system this information so that it will "keep these questions in mind" when designing the graph. If you don't know which kinds of questions, you can just put a couple of high-level questions that you think apply to your data.
(3) Entity Types: this has a very high impact on the final quality of the graph. Think of these as the types of entities that you want to extract from your data, e.g. person, place, event, etc
All of the above help construct the knowledge graph so that it is specifically designed for your use-case.
Looking forward to someone adapting this for Obsidian and other similar tools. As a low-effort user of Obsidian I would love to reap the benefits of appropriate knowledge graphs, but don't want to put that much effort into creating one myself.
Cool! But I'm confused on your pricing. The github page says first 100 requests are free but the landing page says to self host if you want to use for free. I signed up and used the dashboard but I don't see a billing section or option to upgrade the account.
What solutions are folks using to solve queries like "How many of these 1000 podcast transcripts have a positive view of Hillary Clinton"? Seems like you would need a way to map reduce and count? And some kind of agent assigner/router on top of it?
But in general we found the best course of action is simply label everything. Because our customers will want those answers and rag won’t really work at the scale of “all podcasts the last 6 months. What is the trend of sentiment Hillary Clinton and what about the top topics and entities mentioned nearby”. So we take a more “brute force” approach :-)
At the moment this repo is designed to handle more RAG-oriented use cases, i.e. that require to recall the "top pieces of information" relevant to a given question/context. In your specific example, right now, FastGraphRAG would select the nodes that represent podcasts that are connected to Hilary Clinton, feed them to an LLM which would then select the ones that are positively associated with her. As a next step, we plan to weight the connections between nodes given the query. This way, PageRank will explore only edges which carry the concept "positively associated with", and only the right podcasts would be selected and returned, without having to ask an LLM to classify them. Note that this is basically a fuzzy join and so it will produce only a "best-effort" answer rather than an exact one.
I don't have a dev answer, but in case its relevant, I've seen commercial services that I imagine are doing something similar on the back end-- ground news is one of them. I wish they had monthly subs for their top tier plan rather than only annual, but it seems like a cool product. I haven't actually used it though.
What feature(s) of the top tier plan do you wish you had? I have no idea how their subs work but have seen a few ads for the product so have a vague idea that it rates news for bias but don’t see how that would involve many different tiers of subs.
It’s been a while since I looked, but unless they changed it, you needed the top tier plan to get a report analyzing the biases of your reading choices and recommending things to balance it out.
HippoRAG is an amazing work and it was a source of inspiration for us as noted in the references. There are a couple of differences:
(1) FastGraphRAG allows the user to make the graph construction opinionated and specialized on a given domain and for a given use-case; this allows to clear out all the noise in the data and yields better results;
(2) Unlike HippoRAG, FastGraphRAG initializes PageRank with a mixture of semantic retrieval and entity extractions;
(3) HippoRAG is the outcome of an academic paper, and we saw the need for a more robust and production-ready implementation. Our repo is fully typed, includes tests, handles retries with Instructor, uses structured outputs, and so on.
Moving forward, we see our implementation diverge from HippoRAG more radically as we start to introduce new mechanisms such as weighted edges and negative PageRank to model repulsors.
Generally speaking RAG comes in the game when it is impractical to use large context windows for three reasons: (1) accuracy drops as you stuff the context windows, (2) currently, context windows do not scale past 1M tokens, and (3) even with caching, moving millions of tokens is wasteful and not viable both in terms of costs and latency.
So we should really compare this to other RAG approaches. If we compare it to vector databases RAG, knowledge graphs have the advantage that they model the connections between datapoints. This is super important when asking questions that requires to reason across multiple pieces of information, i.e. multi-hop reasoning.
Also, the graph construction is essentially an exercise in cleaning data to extract the knowledge. Let me give you a practical example. Let's pretend we're indexing customer tickets for creating an AI assistant. If we were to store the data on the tickets as it is, we would overwhelm the vector database with all the noise coming from the conversational nature of this data. With knowledge graphs, we extract only the relevant entities and relationships and store the distilled knowledge in our graph. At query time, we find the answer over a structured data model that contains only clean information
Just out of interest: why is every python file prefixed with an underscore? I’ve never seen it before. Is it to avoid collisions with package imports? e.g. “types”
It is to mark the package as private (in the sense that for normal usage you shouldn't need it). We are still writing the documentation on how to customize every little bit of the graph construction and querying pipeline, once that is ready we will expose the right tools (and files) for all of that :) For now just go with `from fast_graphrag import GraphRAG` and you should be good to go :)
Yes, feel free to try it out! You can specialize the graph for the codebase use-case by configuring the graph prompt, entity types, and example questions accordingly.
We are building connectors for that, so it will soon :) At the moment we are using python-igraph (which does everything locally) as we wanted to offer something as ready to use as possible.
I'd like to partner to see if a connector to a graph db can be mutually beneficial and provide some value to users. How do I reach out ? NOTE: Im not from Neo4j
That would be awesome, we have a discord you can join and we can talk there (link is in the github repo, message Antonio)
or you can message antonio [at] circlemind.com
Neat, we are doing something similar with cognee, but are letting users define graph ingestion, generation, and retrieval themselves instead of making assumptions: https://github.com/topoteretes/cognee
I think the main bit here is that the knowledge graph is entirely constructed by LLMs. It's not just using a pre-existing knowledge graph. It's creating a knowledge graph on the fly based on your data.
Navigating the graph, on the other hand, is the perfect task for PageRank.
Exactly! Also PageRank is used to navigate the graph and find "missing links" between the concepts selected from the query using semantic search via LLMs (so to be able to find information to answer questions that require multi-hop or complex reasoning in one go).
What you should note as quaint is probably more like the integration of more "symbolic" strategies to NNs in AI.
Past the initial sensation, it is pretty linear that "something good at language" (an interface) be integrated with "something good at information retrieval" (the data). (Still sought what comes next, "something to give reliability to processing".)
It’s not AI, it’s a collection technologies and practices within the domain of AI, symbolic and sub symbolic. Arguably classic search is another technology/approach/algorithm with the domain of AI.
Not who you're replying to, but from my vantage point, marketing folks seem to be pushing LLM products as replacements for traditional search products. I think what the post is proposing makes perfect sense from a technical perspective, though. The utility of LLMs will come down to good old-fashioned product design, leveraging existing concepts, and novel technical innovation rather than just dumping quintillions of dollars into increasingly large models and hardware until it does everything for us.
Wonder why this all - here on HN - is not part of the readme .md which says absolutely nothing about how and why this all would work.
The whole approach to representing the work, including the writing here, screams marketing, and the paid offering is the only thing made absolutely clear about it.
p.s. I absolutely understand why a knowledge graph is essential and THE right approach for RAG, and particularly when vector DBS on their own are subpar. But so do know many others and from the way the repo is presented it absolutely gives no clue why yours is _something_ in respect to other/common-sense graph-RAG-somethings.
You see, there are hundreds of smart people out there who can easily come to conclusion data needs to be presented as knowledge in graph-ontological way and then feed the context with only the relevant subgraph. Like, you could’ve said so much rather than asking .0084 cents or whatever for APIs as the headline of a presumably open repo.
The problem is neither with HN, the statement was not about HN being or not being about startups (though I would not say personally it is "for startups"), neither it was against startups or other starting groups showing projects. The problem cited is about lack of clarity and supposed change of topic, not the chosen audience.
Now, what is your comment precisely about, cause I'm pretty sure what mine was?
I completely agree that the README could do a better job explaining the implementation details and our reasoning behind key design choices. For instance, we should elaborate on why we believe using PageRank offers a more effective exploration strategy compared to other GraphRAG approaches.
FastGraphRAG is entirely free to use, even for commercial applications, and we’re happy to make it accessible to everyone. The managed service is how we sustain our business.
LLMs are only used to construct the graph, to navigate it we use an algorithmic approach. As of now, what we do is very similar to HippoRAG (https://github.com/OSU-NLP-Group/HippoRAG), their paper can give a good overview on how things are working under the hood!
https://arxiv.org/abs/2105.00110
The paper shows high efficiency compared to other centralities like PageRank, however in some research using the GraphBLAS I and my coauthors found that TC was slower on a variety of sparse graphs than our sparse formulation of PR for graphs up to 1.8 billion edges, but that TC appears to scale better as graphs get larger and is likely more efficient in the trillion edge realm.
https://fossies.org/linux/SuiteSparse/GraphBLAS/Doc/The_Grap...
reply