Introduction
Many generative AI use circumstances nonetheless revolve round Retrieval Augmented Era (RAG), but constantly fall wanting person expectations. Regardless of the rising physique of analysis on RAG enhancements and even including Brokers into the method, many options nonetheless fail to return exhaustive outcomes, miss data that’s essential however sometimes talked about within the paperwork, require a number of search iterations, and customarily wrestle to reconcile key themes throughout a number of paperwork. To high all of it off, many implementations nonetheless depend on cramming as a lot “related” data as attainable into the mannequin’s context window alongside detailed system and person prompts. Reconciling all this data usually exceeds the mannequin’s cognitive capability and compromises response high quality and consistency.
That is the place our Agentic Information Distillation + Pyramid Search Method comes into play. As an alternative of chasing the most effective chunking technique, retrieval algorithm, or inference-time reasoning methodology, my group, Jim Brown, Mason Sawtell, Sandi Besen, and I, take an agentic method to doc ingestion.
We leverage the complete functionality of the mannequin at ingestion time to focus solely on distilling and preserving essentially the most significant data from the doc dataset. This essentially simplifies the RAG course of by permitting the mannequin to direct its reasoning skills towards addressing the person/system directions slightly than struggling to know formatting and disparate data throughout doc chunks.
We particularly goal high-value questions which can be usually tough to judge as a result of they’ve a number of appropriate solutions or answer paths. These circumstances are the place conventional RAG options wrestle most and present RAG analysis datasets are largely inadequate for testing this drawback house. For our analysis implementation, we downloaded annual and quarterly reviews from the final 12 months for the 30 corporations within the DOW Jones Industrial Common. These paperwork could be discovered by means of the SEC EDGAR website. The information on EDGAR is accessible and able to be downloaded for free or could be queried by means of EDGAR public searches. See the SEC privacy policy for added particulars, data on the SEC web site is “thought-about public data and could also be copied or additional distributed by customers of the site with out the SEC’s permission”. We chosen this dataset for 2 key causes: first, it falls exterior the information cutoff for the fashions evaluated, making certain that the fashions can’t reply to questions based mostly on their information from pre-training; second, it’s a detailed approximation for real-world enterprise issues whereas permitting us to debate and share our findings utilizing publicly out there information.
Whereas typical RAG options excel at factual retrieval the place the reply is definitely recognized within the doc dataset (e.g., “When did Apple’s annual shareholder’s assembly happen?”), they wrestle with nuanced questions that require a deeper understanding of ideas throughout paperwork (e.g., “Which of the DOW corporations has essentially the most promising AI technique?”). Our Agentic Information Distillation + Pyramid Search Method addresses a lot of these questions with a lot better success in comparison with different normal approaches we examined and overcomes limitations related to utilizing information graphs in RAG techniques.
On this article, we’ll cowl how our information distillation course of works, key advantages of this method, examples, and an open dialogue on the easiest way to judge a lot of these techniques the place, in lots of circumstances, there is no such thing as a singular “proper” reply.
Constructing the pyramid: How Agentic Information Distillation works
Overview
Our information distillation course of creates a multi-tiered pyramid of data from the uncooked supply paperwork. Our method is impressed by the pyramids utilized in deep studying laptop vision-based duties, which permit a mannequin to investigate a picture at a number of scales. We take the contents of the uncooked doc, convert it to markdown, and distill the content material into an inventory of atomic insights, associated ideas, doc abstracts, and basic recollections/reminiscences. Throughout retrieval it’s attainable to entry all or any ranges of the pyramid to answer the person request.
Easy methods to distill paperwork and construct the pyramid:
- Convert paperwork to Markdown: Convert all uncooked supply paperwork to Markdown. We’ve discovered fashions course of markdown finest for this activity in comparison with different codecs like JSON and it’s extra token environment friendly. We used Azure Doc Intelligence to generate the markdown for every web page of the doc, however there are a lot of different open-source libraries like MarkItDown which do the identical factor. Our dataset included 331 paperwork and 16,601 pages.
- Extract atomic insights from every web page: We course of paperwork utilizing a two-page sliding window, which permits every web page to be analyzed twice. This offers the agent the chance to appropriate any potential errors when processing the web page initially. We instruct the mannequin to create a numbered record of insights that grows because it processes the pages within the doc. The agent can overwrite insights from the earlier web page in the event that they have been incorrect because it sees every web page twice. We instruct the mannequin to extract insights in easy sentences following the subject-verb-object (SVO) format and to write down sentences as if English is the second language of the person. This considerably improves efficiency by encouraging readability and precision. Rolling over every web page a number of occasions and utilizing the SVO format additionally solves the disambiguation drawback, which is a large problem for information graphs. The perception era step can be significantly useful for extracting data from tables for the reason that mannequin captures the details from the desk in clear, succinct sentences. Our dataset produced 216,931 whole insights, about 13 insights per web page and 655 insights per doc.
- Distilling ideas from insights: From the detailed record of insights, we establish higher-level ideas that join associated details about the doc. This step considerably reduces noise and redundant data within the doc whereas preserving important data and themes. Our dataset produced 14,824 whole ideas, about 1 idea per web page and 45 ideas per doc.
- Creating abstracts from ideas: Given the insights and ideas within the doc, the LLM writes an summary that seems each higher than any summary a human would write and extra information-dense than any summary current within the unique doc. The LLM generated summary gives extremely complete information in regards to the doc with a small token density that carries a big quantity of data. We produce one summary per doc, 331 whole.
- Storing recollections/reminiscences throughout paperwork: On the high of the pyramid we retailer essential data that’s helpful throughout all duties. This may be data that the person shares in regards to the activity or data the agent learns in regards to the dataset over time by researching and responding to duties. For instance, we will retailer the present 30 corporations within the DOW as a recollection since this record is completely different from the 30 corporations within the DOW on the time of the mannequin’s information cutoff. As we conduct an increasing number of analysis duties, we will repeatedly enhance our recollections and preserve an audit path of which paperwork these recollections originated from. For instance, we will maintain observe of AI methods throughout corporations, the place corporations are making main investments, and so forth. These high-level connections are tremendous necessary since they reveal relationships and data that aren’t obvious in a single web page or doc.

We retailer the textual content and embeddings for every layer of the pyramid (pages and up) in Azure PostgreSQL. We initially used Azure AI Search, however switched to PostgreSQL for value causes. This required us to write down our personal hybrid search perform since PostgreSQL doesn’t but natively help this function. This implementation would work with any vector database or vector index of your selecting. The important thing requirement is to retailer and effectively retrieve each textual content and vector embeddings at any degree of the pyramid.
This method primarily creates the essence of a information graph, however shops data in pure language, the way in which an LLM natively desires to work together with it, and is extra environment friendly on token retrieval. We additionally let the LLM decide the phrases used to categorize every degree of the pyramid, this appeared to let the mannequin determine for itself the easiest way to explain and differentiate between the data saved at every degree. For instance, the LLM most popular “insights” to “details” because the label for the primary degree of distilled information. Our purpose in doing this was to higher perceive how an LLM thinks in regards to the course of by letting it determine methods to retailer and group associated data.
Utilizing the pyramid: The way it works with RAG & Brokers
At inference time, each conventional RAG and agentic approaches profit from the pre-processed, distilled data ingested in our information pyramid. The pyramid construction permits for environment friendly retrieval in each the standard RAG case, the place solely the highest X associated items of data are retrieved or within the Agentic case, the place the Agent iteratively plans, retrieves, and evaluates data earlier than returning a remaining response.
The good thing about the pyramid method is that data at any and all ranges of the pyramid can be utilized throughout inference. For our implementation, we used PydanticAI to create a search agent that takes within the person request, generates search phrases, explores concepts associated to the request, and retains observe of data related to the request. As soon as the search agent determines there’s enough data to handle the person request, the outcomes are re-ranked and despatched again to the LLM to generate a remaining reply. Our implementation permits a search agent to traverse the data within the pyramid because it gathers particulars a few idea/search time period. That is just like strolling a information graph, however in a means that’s extra pure for the LLM since all the data within the pyramid is saved in pure language.
Relying on the use case, the Agent may entry data in any respect ranges of the pyramid or solely at particular ranges (e.g. solely retrieve data from the ideas). For our experiments, we didn’t retrieve uncooked page-level information since we needed to concentrate on token effectivity and located the LLM-generated data for the insights, ideas, abstracts, and recollections was enough for finishing our duties. In idea, the Agent may even have entry to the web page information; this would offer further alternatives for the agent to re-examine the unique doc textual content; nevertheless, it could additionally considerably improve the entire tokens used.
Here’s a high-level visualization of our Agentic method to responding to person requests:

Outcomes from the pyramid: Actual-world examples
To judge the effectiveness of our method, we examined it in opposition to quite a lot of query classes, together with typical fact-finding questions and sophisticated cross-document analysis and evaluation duties.
Reality-finding (spear fishing):
These duties require figuring out particular data or details which can be buried in a doc. These are the varieties of questions typical RAG options goal however usually require many searches and devour a lot of tokens to reply appropriately.
Instance activity: “What was IBM’s whole income within the newest monetary reporting?”
Instance response utilizing pyramid method: “IBM’s whole income for the third quarter of 2024 was $14.968 billion [ibm-10q-q3-2024.pdf, pg. 4]

This result’s appropriate (human-validated) and was generated utilizing solely 9,994 whole tokens, with 1,240 tokens within the generated remaining response.
Complicated analysis and evaluation:
These duties contain researching and understanding a number of ideas to realize a broader understanding of the paperwork and make inferences and knowledgeable assumptions based mostly on the gathered details.
Instance activity: “Analyze the investments Microsoft and NVIDIA are making in AI and the way they’re positioning themselves available in the market. The report must be clearly formatted.”
Instance response:

The result’s a complete report that executed shortly and accommodates detailed details about every of the businesses. 26,802 whole tokens have been used to analysis and reply to the request with a big share of them used for the ultimate response (2,893 tokens or ~11%). These outcomes have been additionally reviewed by a human to confirm their validity.

Instance activity: “Create a report on analyzing the dangers disclosed by the assorted monetary corporations within the DOW. Point out which dangers are shared and distinctive.”
Instance response:


Equally, this activity was accomplished in 42.7 seconds and used 31,685 whole tokens, with 3,116 tokens used to generate the ultimate report.

These outcomes for each fact-finding and sophisticated evaluation duties show that the pyramid method effectively creates detailed reviews with low latency utilizing a minimal quantity of tokens. The tokens used for the duties carry dense that means with little noise permitting for high-quality, thorough responses throughout duties.
Advantages of the pyramid: Why use it?
Total, we discovered that our pyramid method supplied a big increase in response high quality and general efficiency for high-value questions.
A few of the key advantages we noticed embody:
- Diminished mannequin’s cognitive load: When the agent receives the person activity, it retrieves pre-processed, distilled data slightly than the uncooked, inconsistently formatted, disparate doc chunks. This essentially improves the retrieval course of for the reason that mannequin doesn’t waste its cognitive capability on attempting to interrupt down the web page/chunk textual content for the primary time.
- Superior desk processing: By breaking down desk data and storing it in concise however descriptive sentences, the pyramid method makes it simpler to retrieve related data at inference time by means of pure language queries. This was significantly necessary for our dataset since monetary reviews include a lot of essential data in tables.
- Improved response high quality to many varieties of requests: The pyramid permits extra complete context-aware responses to each exact, fact-finding questions and broad evaluation based mostly duties that contain many themes throughout quite a few paperwork.
- Preservation of essential context: Because the distillation course of identifies and retains observe of key details, necessary data which may seem solely as soon as within the doc is less complicated to keep up. For instance, noting that each one tables are represented in hundreds of thousands of {dollars} or in a specific forex. Conventional chunking strategies usually trigger this kind of data to slide by means of the cracks.
- Optimized token utilization, reminiscence, and velocity: By distilling data at ingestion time, we considerably scale back the variety of tokens required throughout inference, are capable of maximize the worth of data put within the context window, and enhance reminiscence use.
- Scalability: Many options wrestle to carry out as the dimensions of the doc dataset grows. This method gives a way more environment friendly solution to handle a big quantity of textual content by solely preserving essential data. This additionally permits for a extra environment friendly use of the LLMs context window by solely sending it helpful, clear data.
- Environment friendly idea exploration: The pyramid permits the agent to discover associated data just like navigating a information graph, however doesn’t require ever producing or sustaining relationships within the graph. The agent can use pure language solely and maintain observe of necessary details associated to the ideas it’s exploring in a extremely token-efficient and fluid means.
- Emergent dataset understanding: An sudden advantage of this method emerged throughout our testing. When asking questions like “what are you able to inform me about this dataset?” or “what varieties of questions can I ask?”, the system is ready to reply and counsel productive search subjects as a result of it has a extra strong understanding of the dataset context by accessing greater ranges within the pyramid just like the abstracts and recollections.
Past the pyramid: Analysis challenges & future instructions
Challenges
Whereas the outcomes we’ve noticed when utilizing the pyramid search method have been nothing wanting superb, discovering methods to ascertain significant metrics to judge your complete system each at ingestion time and through data retrieval is difficult. Conventional RAG and Agent analysis frameworks usually fail to handle nuanced questions and analytical responses the place many alternative responses are legitimate.
Our group plans to write down a analysis paper on this method sooner or later, and we’re open to any ideas and suggestions from the group, particularly in terms of analysis metrics. Most of the present datasets we discovered have been centered on evaluating RAG use circumstances inside one doc or exact data retrieval throughout a number of paperwork slightly than strong idea and theme evaluation throughout paperwork and domains.
The primary use circumstances we’re concerned with relate to broader questions which can be consultant of how companies truly need to work together with GenAI techniques. For instance, “inform me all the things I have to learn about buyer X” or “how do the behaviors of Buyer A and B differ? Which am I extra more likely to have a profitable assembly with?”. A lot of these questions require a deep understanding of data throughout many sources. The solutions to those questions sometimes require an individual to synthesize information from a number of areas of the enterprise and suppose critically about it. In consequence, the solutions to those questions are not often written or saved anyplace which makes it not possible to easily retailer and retrieve them by means of a vector index in a typical RAG course of.
One other consideration is that many real-world use circumstances contain dynamic datasets the place paperwork are constantly being added, edited, and deleted. This makes it tough to judge and observe what a “appropriate” response is for the reason that reply will evolve because the out there data adjustments.
Future instructions
Sooner or later, we imagine that the pyramid method can tackle a few of these challenges by enabling more practical processing of dense paperwork and storing realized data as recollections. Nevertheless, monitoring and evaluating the validity of the recollections over time will probably be essential to the system’s general success and stays a key focus space for our ongoing work.
When making use of this method to organizational information, the pyramid course of may be used to establish and assess discrepancies throughout areas of the enterprise. For instance, importing all of an organization’s gross sales pitch decks may floor the place sure services or products are being positioned inconsistently. It may be used to match insights extracted from numerous line of enterprise information to assist perceive if and the place groups have developed conflicting understandings of subjects or completely different priorities. This software goes past pure data retrieval use circumstances and would permit the pyramid to function an organizational alignment device that helps establish divergences in messaging, terminology, and general communication.
Conclusion: Key takeaways and why the pyramid method issues
The information distillation pyramid method is critical as a result of it leverages the complete energy of the LLM at each ingestion and retrieval time. Our method permits you to retailer dense data in fewer tokens which has the additional benefit of lowering noise within the dataset at inference. Our method additionally runs in a short time and is extremely token environment friendly, we’re capable of generate responses inside seconds, discover probably tons of of searches, and on common use (this consists of all of the search iterations!).
We discover that the LLM is way higher at writing atomic insights as sentences and that these insights successfully distill data from each text-based and tabular information. This distilled data written in pure language could be very straightforward for the LLM to know and navigate at inference because it doesn’t must expend pointless power reasoning about and breaking down doc formatting or filtering by means of noise.
The power to retrieve and combination data at any degree of the pyramid additionally gives vital flexibility to handle quite a lot of question sorts. This method provides promising efficiency for big datasets and permits high-value use circumstances that require nuanced data retrieval and evaluation.
Be aware: The opinions expressed on this article are solely my very own and don’t essentially mirror the views or insurance policies of my employer.
All in favour of discussing additional or collaborating? Attain out on LinkedIn!