Vector Search and Embeddings for Regulatory Filings

Search for “orbital debris mitigation” across the four U.S. space regulatory agencies and you will find FCC filings immediately. The term appears in every orbital debris mitigation plan submitted with a space station application. But you will miss the FAA Part 450 flight safety analyses that address the same concern under different language: “debris casualty area,” “breakup fragment probability,” “reentry survivability.” You will miss ITU coordination requests flagging interference risks from debris-generating events. You will miss NOAA remote sensing license conditions tied to end-of-life disposal.

Same regulatory concern. Four agencies. Four vocabularies. A keyword search finds what you already know to look for. It misses what you do not.

Semantic search for regulatory filings is retrieval by meaning rather than exact keyword match. An embedding model converts each filing into a high-dimensional vector. Filings that mean similar things produce vectors that sit close together in that space. A query vector is compared against the index using cosine similarity or inner product, and the closest filings are returned regardless of whether they share terms with the query. The result is cross-agency, cross-vocabulary retrieval that keyword search cannot replicate.

Why Keyword Search Fails for Regulatory Data

The problem is not that keyword search is bad technology. BM25, the standard sparse retrieval ranking function described by Robertson and Zaragoza in The Probabilistic Relevance Framework: BM25 and Beyond (Foundations and Trends in Information Retrieval, 2009), remains the strongest pure-keyword baseline in information retrieval and is the workhorse inside systems like Elasticsearch and OpenSearch. The problem is that regulatory data violates the assumption keyword search depends on: that documents about the same topic use the same words.

The FCC calls it a space station application and files it in the International Bureau Filing System (IBFS). The FAA calls it a launch license and processes it under 14 CFR Part 450. The ITU calls it a satellite network filing and routes it through the Space Networks Systems (SNS) database under the Radio Regulations. An operator pursuing a LEO broadband constellation needs all three. They are steps in the same mission. But a keyword search for any one term returns results from only one agency. The same orbital mission lives in three different regulatory vocabularies, and keyword search treats them as unrelated.

This is not limited to cross-agency terminology gaps. Within a single agency, language shifts across filing types. An FCC STA modification and an FCC license amendment may address the same spectrum concern using different phrasing. An IBFS application and an ECFS comment on the same docket describe the same proceeding from different angles, the operator’s filing versus the public’s response, with almost no term overlap.

For an analyst tracking regulatory activity across the FCC, FAA, NOAA, and ITU, the full regulatory stack that governs getting a satellite from concept to orbit, keyword search creates systematic blind spots. Not occasional gaps. Structural ones. Every query is limited to the vocabulary the searcher already knows, which means every query misses filings phrased in terms the searcher has not anticipated.

What Is a Vector Embedding

An embedding is a learned mapping from text to a vector of real numbers. The mapping is produced by a neural model trained so that semantically similar inputs produce vectors that are close together under a chosen distance metric, almost always cosine similarity for normalized vectors.

The lineage is well documented. Word-level embeddings began with word2vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014). Contextual sentence and passage embeddings became practical with the transformer architecture introduced in Attention Is All You Need (Vaswani et al., 2017) and the bidirectional pre-training approach of BERT (Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, NAACL 2019). Sentence-BERT (Reimers and Gurevych, Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks, EMNLP 2019) made dense retrieval at scale practical by training a siamese architecture that produces sentence vectors directly comparable by cosine similarity. More recent open-source models follow the same template at larger scale: Microsoft’s E5 family (Wang et al., Text Embeddings by Weakly-Supervised Contrastive Pre-training, 2022) and Nomic AI’s nomic-embed-text v1.5 (Nussbaum et al., Nomic Embed: Training a Reproducible Long Context Text Embedder, 2024) are widely used production-grade models with permissive licenses.

The output of these models is a vector. nomic-embed-text v1.5 produces a 768-dimensional vector per input. OpenAI’s text-embedding-3-large produces a 3,072-dimensional vector. Each dimension is a learned axis of meaning. Higher dimensionality captures more semantic nuance but costs more to store and search.

The intuition is simpler than the math. Think of it as a coordinate system for meaning. Two filings land near each other in that space because they are about similar things, not because they contain the same strings. An FCC filing about Ka-band interference in LEO and an ITU spectrum coordination request about the same frequency range will embed near each other even though one uses FCC terminology and the other uses ITU notation. The model has learned that these texts are about the same thing.

How Vector Indexes Work in Production

Computing cosine similarity between a query vector and every document vector in a corpus is feasible at small scale. At tens of thousands of filings it works. At millions, brute force is too slow for interactive search.

Production systems use approximate nearest neighbor (ANN) indexes. The two dominant index families are:

• HNSW (Hierarchical Navigable Small World), introduced by Malkov and Yashunin in Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs (IEEE TPAMI, 2018). HNSW builds a multi-layer graph that supports logarithmic-time nearest neighbor queries with high recall. It is the default index in pgvector, Qdrant, Weaviate, and Milvus.
• IVF and IVF-PQ (Inverted File with Product Quantization), developed in the FAISS library by Johnson, Douze, and Jegou at Meta AI (Billion-scale similarity search with GPUs, IEEE Transactions on Big Data, 2019). IVF partitions the vector space into Voronoi cells and searches only the cells closest to the query. Product quantization compresses vectors for billion-scale retrieval.

Index choice involves a recall-latency-memory tradeoff that is documented empirically by ANN-Benchmarks, the standard public comparison harness maintained by Aumueller, Bernhardsson, and Faithfull. There is no single best index. There is the index that matches your corpus size, your latency budget, and your recall target.

The vector index sits inside a broader pipeline. Filings are extracted from source systems, chunked into passages, embedded, and inserted into the index alongside their identifiers and metadata. Queries are embedded with the same model, the index returns the top-k nearest vectors, and the corresponding filing IDs are looked up in the primary store.

From “Search For” to “Similar To”

The paradigm shift is the query model. Keyword search asks: “find documents containing these terms.” Semantic search asks: “find documents similar to this one.”

That is a different kind of question, and it unlocks a different class of analysis. An analyst reviewing an FCC STA modification for a mega-constellation operator can ask: show me other modifications with similar characteristics, same frequency bands, similar orbital parameters, comparable entity profiles. The system returns filings that are thematically related, not just terminally matched. The results include ITU filings, FAA authorizations, and FCC applications that the analyst would not have found with any keyword query, because they would not have known which keywords to use.

Near-duplicate detection becomes possible. The same operator appears under different legal names across agencies. SpaceX in FAA records, Space Exploration Technologies Corp. in FCC filings, SPACEEXPLORATION TECHNOLOGIES CORP in ITU submissions. Keyword search treats these as different entities. Semantic search surfaces the connection because filings from the same operator about the same mission cluster together in vector space regardless of the entity name string.

Amendment and proceeding chains are another natural fit. Regulatory activity cascades. An FCC rulemaking like the five-year deorbit rule generates comments, modifications, and downstream filings that span docket boundaries and filing types. Semantic search surfaces the thematic thread connecting those filings, the regulatory conversation, without requiring the searcher to know every docket number in advance.

Cross-agency entity resolution falls out of the same property. When a single launch program touches FAA Part 450 vehicle operator licensing, FCC payload spectrum authorization, NOAA remote sensing licensing under 15 CFR Part 960, and ITU frequency coordination, semantic retrieval collapses those four agency vocabularies into a single cluster around the program.

The shift from “search for” to “similar to” is not incremental. It changes what questions are possible to ask.

Where Semantic Search Breaks

Semantic search solves real problems. It also introduces new ones, and the failure modes matter for anyone building or buying regulatory intelligence tools.

Out-of-vocabulary terms. Domain-specific jargon like NGSO, ESIM, Schedule S, ITU API filing, and 47 CFR section references may be underrepresented in a model’s training data. If the model has not seen enough examples of a term to learn its meaning, it cannot embed it accurately. The embedding for “Schedule S” might land near “schedule” in general rather than near “technical annex to an FCC space station application,” which is what it actually means. Domain specialization in embedding models is an active area of development. Voyage AI’s partnership with Harvey on legal-specific embeddings is one example, and Cohere’s embed-multilingual-v3 explicitly markets legal and financial domain adaptation. General-purpose models still struggle with the long tail of regulatory terminology.

False similarity. Two filings about Ka-band interference embed closely. One concerns LEO-to-GEO interference coordination and the other involves terrestrial fixed-service protection. The regulatory implications are entirely different: different rules, different agencies, different compliance obligations. The embeddings capture topical similarity but miss the regulatory context that determines what a filing actually means for an operator. Semantic proximity is not regulatory equivalence.

Embedding drift. Swapping or updating the embedding model, whether for better performance, lower cost, or domain adaptation, requires re-embedding every document in the corpus. At scale, this is a non-trivial operation. It is not just compute cost. It is the validation burden: confirming that retrieval quality improved rather than regressed across the query patterns that matter to your users. Asymmetric retrieval models like E5 and BGE further require that queries and documents be embedded with the correct prefix or instruction, and mixing model versions in a single index produces silently degraded results.

Chunking sensitivity. A filing chunked at the paragraph level retrieves differently than the same filing chunked at the section level. Chunk too small and you lose context. Chunk too large and your similarity signal gets diluted. Production systems tune chunk size, overlap, and metadata enrichment per document type, with FCC narratives chunked differently than ITU SNS structured fields.

The production answer is hybrid search. Pure vector search underperforms on exact-match queries, the ones where keyword search excels. Filing numbers, docket IDs, entity names, FRNs, callsigns: these are precise identifiers where you want exact string matching, not semantic approximation. Real systems combine dense vector search (embeddings) with sparse keyword search (BM25 or equivalent) and blend the results using techniques like Reciprocal Rank Fusion, formalized by Cormack, Clarke, and Buettcher in Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods (SIGIR 2009). RRF combines ranked lists from multiple retrievers without requiring score calibration between them. This is not a compromise. It is how production retrieval works.

How to Evaluate Embedding Models on Regulatory Data

There is no leaderboard for FCC filing retrieval. There is no leaderboard for ITU SNS retrieval. The benchmarks that exist measure adjacent capabilities, and you have to know which signals transfer and which do not.

General retrieval benchmarks. The Massive Text Embedding Benchmark (MTEB), introduced by Muennighoff et al. (MTEB: Massive Text Embedding Benchmark, EACL 2023), evaluates embedding models across 58 datasets and 8 task types. BEIR (Thakur et al., BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models, NeurIPS 2021) is the standard zero-shot retrieval benchmark across 18 datasets. MS MARCO, released by Microsoft and Bing in 2016, remains the de facto large-scale passage ranking dataset. These benchmarks tell you whether a model has competent general retrieval behavior. They do not tell you whether it can retrieve regulatory filings.

Legal and regulatory benchmarks. The Massive Legal Embedding Benchmark (MLEB) is the most rigorous evaluation framework for legal and regulatory embedding quality. MLEB results demonstrate that models excelling on general text benchmarks can underperform on domain-specific retrieval tasks. The same pattern shows up across domains: FinBen for finance and LegalBench for legal reasoning both found that general capability is necessary but not sufficient for domain performance, as covered in our piece on benchmarking LLMs for domain-specific extraction.

Building a regulatory golden set. The only conclusive evaluation is on your own data. A regulatory retrieval golden set is a list of queries paired with the filing IDs that should appear in the top results. Queries should reflect actual analyst workflows: cross-agency vocabulary queries, near-duplicate entity queries, amendment chain queries, exact identifier queries that test the hybrid system, and adversarial queries that test out-of-vocabulary handling. NVIDIA and others have shown that 100 to 200 labeled examples are sufficient to meaningfully differentiate model performance on specialized retrieval tasks. The standard metrics are Recall@k, Mean Reciprocal Rank (MRR), and Normalized Discounted Cumulative Gain (NDCG@k). Report all three. Each catches a different failure mode.

Embeddings as the Foundation for AI Workflows

Semantic search is not a standalone feature. It is the retrieval layer in Retrieval-Augmented Generation (RAG), the architecture pattern introduced by Lewis et al. in Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (NeurIPS 2020) that grounds AI-generated analysis in actual source data rather than model training data.

The relationship is direct. When an analyst asks a natural-language question about regulatory filings, RAG works in two steps. First, retrieve the relevant filings using semantic search. Then pass those filings to a language model to generate an answer grounded in the retrieved documents. Without good retrieval, RAG hallucinates. The model answers from its training data instead of your filing database. This is exactly the trust problem that makes AI-generated regulatory intelligence dangerous: plausible answers built on fabricated or misattributed data.

Retrieval quality is necessary but not sufficient. Magesh et al. tested production legal RAG tools and found that even with retrieval grounded in verified legal databases, generation hallucinated at 17 to 33 percent rates. Source grounding requires verification after retrieval, as we cover in AI Verification Engineering. Embeddings solve the finding problem. RAG solves the answering problem. Verification ensures both worked.

Not every retrieval result is equally relevant. Confidence scoring quantifies how similar a result actually is to the query and whether that similarity is meaningful. Low-confidence retrievals that slip through degrade answer quality. In production systems, the cosine similarity score from the index is one input; consensus across multiple retrieval paths, agreement between dense and sparse retrievers, and metadata filters provide additional signals.

Reference Architecture

The pipeline is straightforward, even if the engineering is not.

Source filings (FCC IBFS/ECFS, ITU SNS, FAA Part 450, NOAA CRSRA)
        ↓
Extraction (PDF/HTML to structured fields + clean text)
        ↓
Chunking (per-document-type chunk size and overlap)
        ↓
Embedding (dense vectors) + BM25 indexing (sparse terms)
        ↓
Vector index (HNSW or IVF-PQ) + inverted index
        ↓
Hybrid retrieval (dense top-k + sparse top-k → RRF blend)
        ↓
Confidence scoring + provenance attachment
        ↓
Results (filing IDs with citations) or RAG context window

Raw filings are extracted from agency sources. Text is cleaned and chunked. Embeddings are generated alongside BM25 term indexes. Vectors are indexed in HNSW or IVF-PQ. Queries hit both indexes, results are fused with RRF, and confidence scores are attached to every returned record.

Orbit Sentinel indexes tens of thousands of filings across four agencies and uses nomic-embed-text v1.5, an open-source model that runs locally, keeping filing data off third-party APIs. That architectural choice is not incidental. For regulatory data where provenance and data handling matter, running the embedding model on your own infrastructure means filing text never leaves your systems. It is the same principle behind the verifiable analysis thesis: if you cannot trace how a result was produced, you cannot trust it.

What This Means for Regulatory Intelligence

Semantic search over regulatory filings is not a feature. It is infrastructure.

The analogy is legal citator systems. Shepard’s Citations and KeyCite became table stakes for legal research not because they were novel technology, but because they solved a structural problem: verifying that the authority you are relying on is still good law. Semantic retrieval solves an analogous structural problem for regulatory intelligence: finding the filings that matter across agency boundaries, vocabulary differences, and identifier systems that were never designed to interoperate.

The organizations that build regulatory intelligence on semantic retrieval infrastructure, hybrid search, domain-aware embeddings, confidence-scored results, will have a structural advantage over those still running keyword queries across siloed agency databases. Not because the technology is exotic. Because the alternative does not work when the data speaks four different languages about the same things.

Further Reading

• For why retrieval is necessary but not sufficient, read AI Verification Engineering.
• For how to evaluate LLMs on regulatory tasks, read Benchmarking LLMs for Domain-Specific Extraction.
• For the document parsing layer that feeds retrieval, read Document Extraction at Scale.
• For why AI trust matters for regulatory intelligence, read The Trust Problem.
• For a walkthrough of every agency involved in getting a satellite to orbit, read US Space Regulatory Compliance.
• For deeper definitions of the regulatory terms referenced here, explore our Space Regulatory Glossary.