VC Knowledge Management: Making Past Research Findable

VC knowledge management works when an investor can ask about a company or sector and get every relevant prior record they are allowed to see, each marked with its date and whether it is a fact, an opinion or a decision. That takes five things: start with decision records rather than every file, combine exact-name search with semantic search, attach an as-of date and record type to each document, apply permissions before retrieval, and test the system against questions whose answers you already know. The last step matters most, because a search that misses things feels the same as one that works.

Why can't you judge search by using it?

Because you see what came back, never what stayed hidden.

The classic demonstration is a 1985 study by Blair and Maron, which we read through Grossman and Cormack's account in The Sedona Conference Journal. Searchers on a legal matter were told to retrieve at least 75% of the documents responsive to 51 requests, and to keep going until satisfied. They believed they had reached 75%. They had found 20%. Yet 79% of what they retrieved was relevant, so the results looked good.

Treat this as a warning, not a measure of modern tools. It used 1985 keyword search, and Grossman and Cormack point out real weaknesses: the study was not blind, and the searchers had no way to tell what fraction they had found. That second weakness is the lesson. Precision was high and recall was low, and nobody at the keyboard could see the gap.

A partner who types a company name and gets a good memo is in the same position. The memo is real. The note from a call four years ago, filed under the company's old name, may not have surfaced.

What should a fund index first?

Start with the records that explain decisions, not the whole drive. Rationale is the hardest thing to rebuild, and it is usually what people go looking for. You can extend later: a useful first version does not need every file the firm has ever saved.

Two fields do most of the work. As-of date is when the content was true, not when the file was last touched, because a migration can reset every modified date. Company ID is one record that holds every name the company has used: former names, legal entity, product name and domain. Without it, a company that renamed itself after a pass splits into two histories. Our post on structuring an investment committee memo covers what a memo should contain so that it is worth finding later.

Why does search have to handle exact company names?

Because semantic search alone can miss rare names, and most early-stage companies are rare names.

Sciavolino and colleagues at Princeton tested this at EMNLP 2021 using simple questions built from Wikidata facts. On these entity-rich questions, a dense retrieval model found a passage with the answer in its top 20 results 49.7% of the time on average, against 71.2% for BM25, a traditional keyword-style method. On ordinary benchmark questions the order flipped, 80.1% against 64.5%. Dense models did well on common entities and degraded on rarer ones, and BM25 was less sensitive to how common an entity was.

The study is from 2021, runs on Wikipedia, and tests two specific models, so newer embedding models may do better. The practical conclusion holds anyway: do not assume. Run keyword and semantic search together, route company queries through the company ID, and compare the results on your own archive with the test below. Semantic search earns its place on questions like "what did we conclude about usage-based pricing in vertical software?", where the wording in the memo will differ from the question.

How do you stop old facts being read as current?

Label them, and do not rely on ranking newer documents higher.

The best evidence comes from HoH, a benchmark presented at ACL 2025 by Ouyang and colleagues. It holds 96,124 question-answer pairs built from Wikipedia revisions, each with a current and an outdated answer. Two results matter for a fund.

An old passage does more damage than an irrelevant one. With Llama 3.1 70B, the correct passage plus one irrelevant passage produced a fully correct answer 93.24% of the time and a harmful one 4.12% of the time. The correct passage plus one outdated passage produced 83.00% and 10.05%. The current fact was in front of the model in both cases. The authors also found that models given only outdated material were often confidently wrong, while models given nothing tended to show uncertainty.

Favouring recent documents has a cost. Without any time weighting, the search engine's top five held the current passage 87% of the time and the outdated one 88%. With a time-decay function, outdated hits fell to 50%, but current hits fell to 70%. For a fund this trade is worse than it looks, because the old pass note is sometimes exactly the document you need. Burying it by age removes the record you were looking for.

The benchmark uses Wikipedia, not fund documents, and tests answers generated by models rather than investors reading results. The direction is still clear. Age should be shown, not used to hide things, and different kinds of record age differently.

These defaults are our suggestion, not research findings. Adjust them to your fund.

How should permissions work in a fund's archive?

Apply them before retrieval, so restricted text never reaches the model or the ranked list. A filter added afterwards is not enough, because a summary can repeat a restricted fact without showing the file.

OWASP's 2025 entry on vector and embedding weaknesses names this risk: misaligned access controls can let a model retrieve and disclose sensitive content, and context can leak between classes of users who share a vector database. It recommends fine-grained, permission-aware stores and strict logical partitioning of data between groups.

A fund's groups are unusual, so map them before building anything:

• Deal team only: live diligence, and material received under an NDA that limits who may see it.
• Board member only: portfolio company board papers.
• Partners only: personnel matters, compensation and sensitive reference calls.
• Restricted by relationship: material from a company whose competitor the fund now backs.
• Firm-wide: retired theses, closed sector work and approved summaries.

Then handle two events that are easy to miss. When someone leaves or changes role, access must change the same day. When a document is deleted or a retention period ends, it must also leave the index, or the text survives in search after the file is gone. Which obligations apply depends on your NDAs, board roles and policies, so confirm the tiers with counsel.

How do you test whether your search works?

Write 30 questions with known answers, run them, and score what was missed. This gives a fund something it rarely has: a measure of recall on its own documents.

1. Ask two long-tenured people to list the answers first. For each question, name every record that should appear. This is what the Blair and Maron searchers never had.
2. Write ten company questions. Include at least three companies that changed name, and three from your oldest records.
3. Write ten concept questions in words the original documents did not use, such as a sector theme or a reason for passing.
4. Write five staleness questions where the archive holds a newer and an older answer. Check that both appear with dates and that the older one is labeled.
5. Write five permission questions. Run them from a junior account. The right result is nothing, and no hint in any summary.
6. Score each question. Record how many listed records appeared in the top 10, and which were missed.
7. Fix the causes, not the score. A miss usually traces to a missing alias, an unindexed folder, a scanned PDF with no text layer or absent metadata.
8. Repeat each quarter, and add a new question whenever someone says they could not find something.

We know of no benchmark for a good recall score in a venture archive, so use your first result as a baseline. Permission tests are different: a single failure should stop rollout until it is fixed.

Will better retrieval improve decisions?

Not by itself, and it can hurt if prior work replaces fresh thinking.

Haas and Hansen studied 182 sales proposals at one consulting firm, published in Strategic Management Journal in 2005. Highly experienced teams were more likely than inexperienced teams to lose bids when they used the firm's documents and colleagues' advice. The authors link this partly to relying on "insufficiently tailored or out-of-date prior work". There were also situations where using firm knowledge helped, so value depended on the task and the team's need to learn. It is one firm, in a setting where use of knowledge was strongly encouraged, and winning a bid is not an investment outcome.

The lesson for a fund: use the archive to ask better questions, not to fill in the new memo. Show the prior pass reason and ask what has changed. Do not reward people for how often they cite old work.

Where can AI help, and where should it stop?

AI suits the retrieval and preparation: resolving company names, finding related records, summarizing them with citations and flagging what is old or contradictory. When records disagree, it should show the disagreement rather than invent a consensus. A person should open the source before relying on it, and whether a past view still holds is a judgment for the investment team. For where this sits among a firm's other systems, see our guide to the AI-first VC operations stack.

Permissions, record types and staleness rules differ by fund, which is why Sonexa Labs treats this as commissioned work rather than a packaged product. Through Institutional Knowledge and Firm Operations, one of our Tailored Solutions, a system could retrieve prior memos and notes from repositories your fund approves, respect the access rules you define and flag time-sensitive context, with your team checking sources before using prior material.