Internal Intelligence.

Internal intelligence at a VC firm comes in three stages. They are not alternatives as much as maturity curves. Most firms are at zero. Some are at one. A handful are at two. Almost nobody is at three.

• Stage one — the AI's eyes begin to open. The LLM is connected to your systems of record and data vendors. It can read and write. It knows where things live and can access them with little to moderate prompting. Slower, with no advanced vision, but a great start.
• Stage two — the AI's eyes are wide open but it's frozen. The entire corpus of data is indexed into a vector store. The LLM can retrieve the right twenty five individual datapoints from anywhere in the firm to synthesize the answer a hard question on demand, quickly. It sees everything, but it does not yet adapt and improve over time.
• Stage three — the AI's eyes are open and it can move freely. The system learns iteratively. It improves as the firm uses it. This is aspirational — even the most AI native firms do not have two feet in this stage.

Stage one is connecting every system you already have — CRM, productivity suite, file drive, data providers, vendors — to a central LLM. The LLM becomes the surface where the firm asks questions and does work, and the underlying tools become the hands and eyes.

This is the difference between an LLM out of the box and one that has context. Out of the box, it knows nothing about your firm, your portfolio, your markets, or the company you met with last Tuesday. Connected to your systems, it can answer questions like "summarize every call we've had with this founder," "which companies in our pipeline matched the thesis we wrote in Q2," "draft a follow-up to the partner who flagged this deal."

It is easier than people think. Stage one mostly means connecting tools natively, via MCP, or via traditional APIs when those fail. The LLM ecosystem has converged hard on this — Claude and ChatGPT both have mature connector stories, and MCP has become the dominant protocol for surfacing tools to a model. A reasonably organized firm can get to stage one in just 

Structure it with a skill. If your LLM connects to a file drive, it is worth writing a skill that tells the connector exactly how the drive is organized — where company folders live, how naming works, what the standard subfolders are. Without that context the model resets on every query, burns tokens rediscovering the layout, and still misses the right documents. A single structured instruction removes that friction and turns the drive from a blind search into a directed lookup. We publish a template for exactly this in the File Drive Connector Config skill.

Why it matters. Increasingly, the LLM is where work happens. Systems like Claude have crossed the line from "useful assistant" to "place I do my work." Stage one is what makes that real for a firm — every team member starts their day inside a system that knows the firm cold.

Stage two is RAG — retrieval augmented generation — over the entire body of firm knowledge. Every document, every call transcript, every memo, every deck, every CRM record, indexed into a vector store and queryable by meaning, not just keyword.

The shape: an ingestion pipeline takes everything in your systems, chunks it, embeds it, and stores the vectors. When a user asks a question, the system retrieves the most semantically relevant chunks — the twenty five paragraphs out of thousands or even hundreds of thousands of documents that actually matter — and feeds them to an LLM as context. The answer is grounded, cited, and pulls from places no human would have thought to look.

Our pick: LlamaIndex. LlamaIndex is what we use and what we recommend. It is at the frontier of document ingestion, it is flexible enough to handle the messy variety of file types a firm actually has, and the team is shipping against the right problems — including agentic document ingestion, which is the next leap. Budget roughly $25,000–$30,000 annually for a real deployment. It is not cheap. It is also one of the highest- leverage line items in an AI-native firm.

Alternatives worth knowing. Haystack (from deepset) is a credible open-source alternative — a Python framework for composing RAG and agentic pipelines with more explicit control over each step, and a managed enterprise tier if you do not want to run it yourself. Worth a look if your team has the engineering appetite. Two adjacent options sit further out: Elastic, if you already run Elasticsearch and want to layer vector search onto an existing index rather than stand up a new store; and Glean, which packages indexing, permissions, and a chat surface into a turnkey enterprise product — overkill and overpriced for most firms, but a real option for larger shops that want to buy rather than build.

Orchestration is what makes it work. A vector index over just your file drive is useful. A vector index that also pulls in every CRM note, every call recording transcript, every research report, every newsletter — synced continuously into a single source of truth — is transformative. This is why the orchestration blueprint matters so much for stage two: every workflow that pipes data into your central store makes the brain meaningfully smarter.

Organization is the multiplier. The more organized your underlying systems are, the better RAG works. A file drive where every company has a folder with standard subfolders is dramatically easier to retrieve from than a sprawl of "Untitled deck (3).pptx." The systems of record blueprint covers this in detail — it is the foundation stage two sits on.

The BEAST (sounds cool, right?) is General Innovation's stage 2 system for internal intelligence. Don't ask us what it stands for .. business, enterprise, something something ..

Single source of truth. All firm knowledge ultimately flows into SharePoint. Content originates wherever it originates — CRM, call notes, emails, data rooms, third party research — and our automations centralize it consistently. Today the BEAST is trained on roughly 30,000 documents spanning deal specific diligence and broader market research. The more we capture early in diligence, the more useful it becomes.

The semantic layer. LlamaIndex sits at the core. It ingests almost any document and organizes it into a large database that can be searched by meaning rather than exact words. This is what lets the BEAST find the right paragraph in a 200 page data room even when the question is phrased completely differently from the source.

The retrieval flow. When a user asks a question, the BEAST rewrites it to optimize for semantic search, stripping out instructions irrelevant to retrieval. It runs the optimized query across the corpus and pulls back the most meaningful segments, then reviews each segment with an AI model that extracts only the information directly relevant to the question and discards the rest. What reaches the answer model is a tightly curated body of context — file names, links, relevance scores.

The deployment. The retrieval logic lives in a custom application of roughly 500 lines of code, built on GitHub and accessed through Railway, exposed through an API connected to ChatGPT as a custom action. When a question is asked, ChatGPT sends it to the BEAST, receives the curated context, and generates the answer. The ranked, cited structure keeps answers grounded in real sources and minimizes hallucination.

Beyond the chatbot. Custom instructions shape how the BEAST reasons and responds — emphasizing connection across sources and helping the team build conviction. The BEAST is also integrated into memo generation and other diligence tools, functioning not as a standalone chatbot but as a reusable component inside broader AI workflows that continue to expand.

Google Workspace Intelligence sits somewhere between stage one and stage two. Because the Google ecosystem is unusually tight — Gmail, Drive, Docs, Meet, Calendar all speak to each other natively — their cross-product retrieval tool gets surprisingly close to a real RAG experience without the firm having to build anything.

If you are deciding between Google and Microsoft for the productivity suite, this is a real factor in Google's favor. Microsoft Copilot is not in the same league for this use case. It is one of the strongest practical arguments for choosing Google as the backbone of the firm.

That said, it is not a substitute for stage two. The ceiling on what Workspace intelligence can do is well below what a purpose-built RAG system on top of LlamaIndex can do — both on retrieval quality and on how deeply it can plug into the rest of your workflow stack. We have heard mixed reviews from firms that have tried it, and the reality is less exciting than the demo if you have never used it before.

Stage three is a system that improves iteratively as the firm uses it. The eyes are open, and now it can walk. It thinks. It adjusts. Almost nobody has built this — including us — but firms are beginning to attempt it and it is worth naming so you can see the full arc.

Two directions we find interesting:

• An LLM wiki, à la Karpathy. A living documentation layer that updates over time and grows to represent how the firm actually thinks. Less a static knowledge base, more an opinionated, evolving brain that new team members can read to absorb the firm in a week.
• An investment decision engine. A model that gets better at evaluating opportunities as it is fed more data about how the firm has actually invested — what worked, what didn't, where the partnership's instincts were right, where they were wrong.

We are in the very early innings. This is unproven and we list it more as a horizon than a recommendation. But the firms that get stage three right in the next five years will have an edge that is genuinely hard to compete with.

You do not have to wait to start. These systems take time to develop and they compound — the earlier you plant stage one, the better stage two and stage three will be when you get there. Take the low-hanging fruit first: get the quick wins from stage one, feel the lift, and then expand from there toward a real stage two build.

The order we recommend:

1. Get your systems of record and file drive into a defensible state. Standard folder structure per company. Consistent naming. Real hygiene.
2. Connect those systems to your LLM OS of choice via native connectors, MCP, or APIs. This is stage one. You will feel the lift immediately.
3. Stand up an ingestion pipeline into a vector store and wire it to your LLM as a retrieval tool. This is stage two. Plan for a real budget and a real ongoing investment.
4. Once stage two is humming, start experimenting at the edges of stage three. A wiki that the firm actively maintains alongside the model. Feedback loops on the memos and decisions the system helps produce.