need semantic search for documents with date filters

Yes — you can build semantic search over documents with date filters. The usual pattern is:

1. Ingest documents
2. Extract text + metadata (e.g. created_at, published_at, source, author)
3. Create embeddings for the text
4. Store embeddings + metadata in a vector database or search engine
5. At query time:
   • run semantic similarity search
   • apply date range filters on metadata

Common architectures

1) Vector DB with metadata filtering

Good if you want semantic search first, then filter by dates.

Examples:

• Pinecone
• Weaviate
• Qdrant
• Milvus
• Chroma (smaller setups)

Typical query:

• similarity: “find documents about contract renewal”
• filter: date >= 2024-01-01 AND date <= 2024-12-31

2) Search engine + vectors

Good if you also need strong keyword search and faceting.

Examples:

• Elasticsearch / OpenSearch
• Azure AI Search

These support:

• vector similarity
• text search
• structured filters like dates

3) Hybrid search

Best for quality in many cases:

• semantic/vector ranking
• keyword match
• date filters
• reranking

---

Recommended data model

Store each document chunk with metadata like:

{
  "id": "doc_123_chunk_4",
  "text": "....",
  "embedding": [0.12, -0.07, ...],
  "metadata": {
    "document_id": "doc_123",
    "title": "Quarterly report",
    "created_at": "2025-01-15",
    "updated_at": "2025-01-20",
    "source": "sharepoint"
  }
}

If documents are long, chunk them and keep the document date on every chunk.

---

Example query logic

User asks:

“Find documents about vendor risk from last quarter”

Translate to:

• semantic query: "vendor risk"
• date filter: last quarter range

Example:

• 2025-04-01 to 2025-06-30

Then search only documents whose metadata date falls in that range.

---

Important design choices

Which date to filter on?

Decide whether to use:

• created_at
• published_at
• modified_at
• document_date (business date)

This matters a lot. Usually you should expose one canonical date field for filtering.

Filter before or after semantic search?

Usually:

• apply date filter at retrieval time if your database supports it
• if not, retrieve top-N semantically then post-filter, but that can hurt recall

Chunk-level or document-level filtering?

If documents are chunked:

• keep metadata on each chunk
• filter chunks by document date

---

If you want a simple stack

Small/medium scale

• Python
• embeddings from OpenAI / sentence-transformers
• Qdrant or Chroma
• date metadata filters

Enterprise/search-heavy

• OpenSearch or Elasticsearch
• vectors + BM25 + date facets

---

Example with Qdrant-style filter

Conceptually:

search(
    query_vector=embedding("vendor risk"),
    filter={
        "must": [
            {"field": "created_at", "gte": "2025-04-01"},
            {"field": "created_at", "lte": "2025-06-30"}
        ]
    }
)

---

If you want, I can help with any of these next:

1. a system design
2. a Python implementation
3. choosing between Pinecone / Qdrant / Elasticsearch
4. adding natural-language date parsing like “last month”, “Q2 2025”

If you tell me your stack and document volume, I can recommend the best approach.