fastapi-semcache 0.4.0

Ultra-lightweight semantic caching middleware for FastAPI APIs and LLM endpoints

fastapi-semcache

Ultra-lightweight semantic caching middleware for FastAPI APIs and LLM endpoints.

fastapi-semcache adds semantic response caching as a thin async middleware layer. Vector similarity search runs inside Postgres via pgvector -- Python never owns the heavy computation. It works as FastAPI middleware today and can also run as a reverse proxy in front of an upstream API or LLM service.

The PyPI distribution and GitHub repository are fastapi-semcache. The import package remains semanticcache (fastapi_semcache is available as an alias).

Why fastapi-semcache?

fastapi-semcache is an ultra-lightweight middleware with a minimal, honest dependency footprint. The core install adds four packages to your project: fastapi, pydantic-settings, psycopg (libpq C bindings), and httpx. No vector database service to run, no ML runtime in the hot path, no framework to build around.

Python is the glue, not the bottleneck. In the request hot path, Python parses the JSON body, dispatches async I/O, and coordinates the result. That is all. Every computationally heavy operation is offloaded:

What	Where it runs
Cosine / ANN vector similarity	Postgres + pgvector (C, indexed)
Embedding generation	Your provider's API (I/O, not CPU)
Response blob storage and retrieval	Postgres rows or Redis (C clients)
HTTP proxying	httpx.AsyncClient (async I/O)

Because all meaningful work is either I/O-bound (GIL released) or executing inside a C extension, Python is never the ceiling even under high concurrency with a single uvicorn worker. The likely bottlenecks under load are your Postgres connection pool and your embedder API, not this middleware.

It plugs into FastAPI with minimal refactoring, while giving you direct control over embeddings, similarity thresholds, vector storage, and cache behavior. The default setup keeps things simple: find the highest-similarity match, apply a threshold, and return a cached response only when it is safe to do so.

It supports FastAPI middleware as a first-class integration path and can also run as a reverse proxy in front of an upstream API or LLM service. Planned support for Django and Flask will extend the same integration model to other Python web stacks.

When not to use

Semantic caching works best when a meaningful fraction of queries share similar intent within your cache window. It is not a good fit for every endpoint:

• Mostly unique queries: creative writing, one-off analysis, ad-hoc exploration. Low hit rates mean you pay embedding cost on every request for little return.
• Highly personalised responses: where the same question requires a different answer per user. Scope partitioning helps, but if every partition is unique you gain nothing.
• Constantly changing data: real-time prices, live feeds, rapidly-evolving documents. Cached responses go stale faster than they help.
• Strict correctness requirements: legal, medical, financial contexts where a semantically similar but not identical query may need a materially different answer.
• Side-effectful POST endpoints: returning a cached response silently skips the side effect. Exclude those paths with a custom extract_query that returns None.
• Exact-match inputs: if your prompts are deterministic and identical across users, a plain Redis key is simpler and cheaper than a vector search.

See docs/when-not-to-use.md for a fuller treatment with alternatives.

Install

pip install fastapi-semcache

Custom embedders: subclass BaseEmbedder from semanticcache.embedders and pass it to SemanticCache(embedder=...) to skip the optional embedding extras. See docs/embedders.md.

Optional extras:

• redis: Async Redis client (redis>=7.4.0) for TTL-backed response blobs when SEMANTIC_CACHE_REDIS_URI is set. Core installs omit it so Postgres-only deployments avoid pulling Redis.
• embed-huggingface: Sentence Transformers and PyTorch. Default PyPI wheels are CPU; for CUDA, install with PyTorch's --extra-index-url (below).
• embed-openai: OpenAI embeddings (openai, tiktoken).
• embed-voyage: Voyage AI embeddings (voyageai, aiohttp).
• embed-ollama: Ollama embeddings via the OpenAI-compatible HTTP API (openai only).

Dependency notes:

• Core fastapi-semcache has no LangChain dependency.
• Core does not include the redis PyPI package; use pip install "fastapi-semcache[redis]" whenever you configure a non-empty Redis URI (otherwise the first Redis use raises ImportError with an install hint).
• Optional extras only add their listed packages (redis, sentence-transformers/torch, openai/tiktoken, voyageai/aiohttp, or openai alone for embed-ollama).

Hugging Face / Sentence Transformers

pip install "fastapi-semcache[embed-huggingface]"

That pulls CPU PyTorch from PyPI. For GPU (CUDA), use the same extra but pass PyTorch's wheel index so pip resolves CUDA builds. Pick a CUDA version that matches your system from PyTorch Get Started:

pip install "fastapi-semcache[embed-huggingface]" \
  --extra-index-url https://download.pytorch.org/whl/cu124

OpenAI embeddings

Install the OpenAI extra so embedder_type="openai" works (pulls openai and tiktoken). Set OPENAI_API_KEY in your environment.

pip install "fastapi-semcache[embed-openai]"

Voyage embeddings

Install the Voyage extra so embedder_type="voyage" works (pulls voyageai and aiohttp). Set VOYAGE_API_KEY or SEMANTIC_CACHE_VOYAGE_API_KEY. Optional SEMANTIC_CACHE_VOYAGE_EMBEDDING_MODEL and SEMANTIC_CACHE_VOYAGE_EMBEDDING_DIMENSIONS default to voyage-4 and 1024 when unset (they must match your chosen model and pgvector column width). Set SEMANTIC_CACHE_VOYAGE_INPUT_TYPE to query or document when you want Voyage’s input-type hint on each request.

pip install "fastapi-semcache[embed-voyage]"

Ollama embeddings

Install the Ollama extra so embedder_type="ollama" works (pulls openai only). Set SEMANTIC_CACHE_OLLAMA_EMBEDDING_MODEL and SEMANTIC_CACHE_OLLAMA_EMBEDDING_DIMENSIONS to match the embedding model you run (dimensions must match pgvector). Optionally set SEMANTIC_CACHE_OLLAMA_BASE_URL (default http://127.0.0.1:11434/v1) and OLLAMA_API_KEY or SEMANTIC_CACHE_OLLAMA_API_KEY when your server uses auth.

pip install "fastapi-semcache[embed-ollama]"

Redis response cache

Install the Redis extra when SEMANTIC_CACHE_REDIS_URI (or constructor redis_uri) is non-empty so redis.asyncio is available.

pip install "fastapi-semcache[redis]"

You can combine extras, for example pip install "fastapi-semcache[redis,embed-openai]" or pip install "fastapi-semcache[redis,embed-voyage]".

FastAPI middleware

Security: cache scope and cross-tenant isolation

By default (SEMANTIC_CACHE_REQUIRE_CACHE_SCOPE=true), the middleware reads the cache partition key from client-controlled sources: the X-Semantic-Cache-Scope header and the cache_scope / tenant_id JSON body fields. Any client can forge these values to read another tenant's cached responses or write responses into another tenant's cache partition.

This default is safe only for single-tenant apps (consider setting SEMANTIC_CACHE_REQUIRE_CACHE_SCOPE=false to remove the scope requirement entirely) or when a trusted edge/gateway overwrites those fields from verified identity before requests reach your app.

For multi-tenant APIs exposed directly to clients, always supply a server-side extract_scope:

from starlette.requests import Request
from semanticcache.middleware.core.extractors import trusted_extract_scope_from_server_side

async def extract_scope(request: Request, body: bytes) -> str | None:
    return await trusted_extract_scope_from_server_side(request)

app.add_middleware(SemanticCacheMiddleware, cache=cache, extract_scope=extract_scope)
# Starlette executes middleware in reverse addition order:
# YourAuthMiddleware runs first and populates request.state,
# then SemanticCacheMiddleware reads it.
app.add_middleware(YourAuthMiddleware)

trusted_extract_scope_from_server_side reads only request.state, which clients cannot forge. See docs/cache-tuning.md for details.

Add SemanticCacheMiddleware to your app and reuse one SemanticCache instance for all requests. Configure Postgres, Redis, and the embedder with SEMANTIC_CACHE_* environment variables (see .env.example). By default only POST requests are intercepted; the middleware derives cache-key text from JSON bodies using query, prompt, input, or chat-style messages (see default_extract_query in semanticcache.middleware). Successful responses whose body parses as a JSON object are candidates for storage, and cache hits replay the original HTTP status and response metadata.

Redis is optional. If SEMANTIC_CACHE_REDIS_URI is empty (or whitespace), the cache runs in Postgres-only mode: semantic lookup and response storage still work via pgvector, but Redis TTL-based payload caching is disabled. If you do set a Redis URI, install fastapi-semcache[redis] (see Redis response cache).

from typing import Any

from fastapi import FastAPI

from semanticcache import SemanticCache, SemanticCacheMiddleware

app = FastAPI()
cache = SemanticCache()
app.add_middleware(SemanticCacheMiddleware, cache=cache)


@app.post("/v1/chat/completions")
async def chat_completions(body: dict[str, Any]) -> dict[str, Any]:
    # Clients should send JSON with prompt, query, input, or chat messages so the
    # middleware can build the cache key (see default_extract_query). By default a
    # tenant scope is also required (header X-Semantic-Cache-Scope or JSON
    # cache_scope / tenant_id); those values are client-controlled unless you replace
    # extract_scope — unsuitable for multi-tenant production without a trusted edge
    # or server-side scope (see docs/cache-tuning.md). Misses run your handler;
    # hits short-circuit with a cached JSON body.
    return {"choices": [{"message": {"role": "assistant", "content": "Hello"}}]}

Run with uvicorn mymodule:app --host 0.0.0.0 --port 8000.

Custom cache key text (extract_query)

If your JSON body does not follow the usual query / prompt / messages patterns, pass an async callable as extract_query. It receives the Starlette Request and the raw body bytes (already buffered by the middleware). Return a non-empty string to embed and look up; return None to skip semantic caching for that request (the route still runs).

If extract_query or extract_model raises, the middleware logs the error (with stack trace) and forwards the request upstream without calling the cache (same outcome as returning None from extract_query, but the route still runs).

You can wrap default_extract_query and add fallbacks for your own fields, or replace it entirely.

from fastapi import FastAPI, Request

from semanticcache import SemanticCache
from semanticcache.middleware import SemanticCacheMiddleware, default_extract_query

async def extract_query(request: Request, body: bytes) -> str | None:
    base = await default_extract_query(request, body)
    if base is not None:
        return base
    # Parse ``body`` for your schema; return None to bypass the cache.
    return None

app = FastAPI()
cache = SemanticCache()
app.add_middleware(
    SemanticCacheMiddleware,
    cache=cache,
    extract_query=extract_query,
)

Use extract_model when the cache key should also vary by model id from headers or JSON (same async (request, body) -> str | None idea). That model id is passed through to SemanticCache.get / put, which scope Postgres rows and Redis payload keys per model bucket as described in docs/cache-tuning.md.

Use extract_scope (optional) when you need custom tenant or user routing; otherwise, with SEMANTIC_CACHE_REQUIRE_CACHE_SCOPE left at its default true, the middleware uses default_extract_scope_from_request_context, which reads X-Semantic-Cache-Scope and JSON cache_scope / tenant_id (numeric tenant_id is accepted). That default is appropriate only for single-tenant deployments or when a trusted gateway overwrites those fields from authenticated identity; otherwise clients can spoof another tenant id and probe or pollute another partition. For multi-tenant APIs exposed to clients, pass extract_scope that derives scope from server-side identity (see trusted_extract_scope_from_server_side in semanticcache.middleware.core.extractors after auth middleware sets request.state). Set SEMANTIC_CACHE_REQUIRE_CACHE_SCOPE=false only for single-tenant apps or isolated cache storage. Scope rules in middleware match SemanticCache.settings when cache is a SemanticCache instance ( cache_settings still drives circuit breaker and flight-lock options). resolve_cache_scope matches the same rules for direct SemanticCache use.

See docs/cache-tuning.md for upgrade notes on scope_key and Redis key layout.

For create_semantic_cache_proxy_app, pass extract_query=... (and other middleware options) as keyword arguments; they are forwarded to SemanticCacheMiddleware.

Use validate_response when a route or provider has a strict response schema and you want to avoid storing malformed payloads. The callback receives a ResponseValidationContext with the request, raw request body, upstream response, parsed JSON object, model, and scope. Return False to return the upstream response normally but skip the cache write.

from semanticcache import ResponseValidationContext


def validate_response(context: ResponseValidationContext) -> bool:
    if context.request.url.path == "/v1/chat/completions":
        return isinstance(context.payload.get("choices"), list)
    return True


app.add_middleware(
    SemanticCacheMiddleware,
    cache=cache,
    validate_response=validate_response,
)

Other advanced options (path_prefix, HTTP 429 circuit breaker via cache_settings, enabled=False) are documented on SemanticCacheMiddleware in semanticcache.middleware.adapters.fastapi (or via the public import semanticcache.middleware). On shutdown, call await cache.close() from a lifespan handler if you want pools closed cleanly. The 429 circuit breaker is per-process; in multi-worker deployments each worker tracks its own counter independently.

Cache behavior and tuning

SemanticCache uses a two-stage retrieval pipeline:

• A primary similarity threshold (SEMANTIC_CACHE_THRESHOLD) and top-k candidate limit (SEMANTIC_CACHE_TOP_K_CANDIDATES) control which nearest neighbors are fetched from pgvector.
• An optional rejection threshold (SEMANTIC_CACHE_REJECTION_THRESHOLD) can then filter out borderline matches; if no candidate passes this second stage, the middleware returns a cache miss.
• Dependency timeouts let you fail fast when providers or storage are slow: SEMANTIC_CACHE_EMBED_TIMEOUT_SECONDS applies to embedder calls, and SEMANTIC_CACHE_STORE_TIMEOUT_SECONDS applies to Postgres/Redis operations. On timeout, SemanticCache raises a timeout error, middleware logs it, and request handling continues in fail-open mode.
• In-flight lock registry cap bounds middleware memory used for concurrent miss coordination: SEMANTIC_CACHE_MIDDLEWARE_FLIGHT_LOCK_MAX_ENTRIES limits retained (query, model, scope) lock keys and evicts least-recently-used unlocked entries when needed.
• Body size limit (SEMANTIC_CACHE_MAX_BODY_BYTES, default 10_485_760 - 10 MB): the middleware skips semantic caching for any request body that exceeds this limit and forwards the request upstream unchanged. Set to 0 to disable the limit entirely. DEFAULT_MAX_BODY_BYTES is exported from semanticcache if you need the constant directly.

See docs/cache-tuning.md for concrete tuning tips and examples.

Reverse proxy

The reverse proxy mode is optional: it forwards traffic to an upstream base URL while using the same semantic cache middleware. Use it when you want a standalone hop in front of another service rather than importing routes into your FastAPI app.

Minimal programmatic setup:

from semanticcache import SemanticCache, create_semantic_cache_proxy_app

cache = SemanticCache()
app = create_semantic_cache_proxy_app(
    upstream="http://127.0.0.1:11434",
    cache=cache,
)

Run with uvicorn mymodule:app --host 0.0.0.0 --port 8080.

This repository includes a small ASGI app at app/main.py (import app for uvicorn). Set SEMANTIC_CACHE_PROXY_UPSTREAM to the backend base URL; the default is http://127.0.0.1:11434. For semantic caching in front of a single trusted upstream, set SEMANTIC_CACHE_REQUIRE_CACHE_SCOPE=false unless you forward a tenant header or JSON scope from clients.

If your upstream requires an Authorization header (for example OpenAI-compatible APIs), set SEMANTIC_CACHE_CACHE_AUTHORIZED_REQUESTS=true or the middleware will bypass cache reads and writes for those requests.

uv run uvicorn app.main:app --host 0.0.0.0 --port 8080

See create_semantic_cache_proxy_app in semanticcache.proxy for timeout, TLS verification, httpx_client_kwargs, and middleware options such as path_prefix and extract_query.

Streaming and chunked responses

By default, fastapi-semcache uses a buffered response mode: the middleware buffers the full downstream response before sending it to the client and before writing to the cache.

Miss delivery: response_mode

Set response_mode on CacheSettings to control how cache-miss responses are delivered:

• "buffered" (default) -- full response is buffered before the client receives anything; cache write completes before the response is returned.
• "tee" -- chunks are forwarded to the client as they arrive; the cache write runs in a background task after the stream completes. This gives lower time-to-first-byte for streaming upstreams (for example token streaming or SSE) while still accumulating the full body for storage.

In both modes the cache stores only fully assembled JSON object responses. The tee path respects the same store rules as the buffered path (headers, validation, and size limits).

Hit delivery: hit_response_mode

Set hit_response_mode on CacheSettings to control how cache-hit responses are delivered:

• "single" (default) -- the cached body is returned as a single HTTP response.
• "stream" -- the cached body is emitted as ASGI body chunks, matching the framing of a streaming miss response. When response_mode="tee" and hit_response_mode is not explicitly set, it defaults to "stream" automatically so hit and miss delivery are symmetric.

Use hit_stream_chunk_size (env SEMANTIC_CACHE_HIT_STREAM_CHUNK_SIZE, default 0) to split hit responses into multiple chunks of at most that many bytes. 0 sends the full body as a single chunk, which is sufficient for most clients; positive values are useful for clients that measure time-to-first-byte or process tokens incrementally.

Reverse proxy note

For create_semantic_cache_proxy_app, upstream responses are fetched via httpx.AsyncClient using a buffered body, but response_mode and hit_response_mode control delivery to clients at the ASGI layer in the same way as the middleware.

Current features

• Huggingface embeddings via Sentence Transformers (embedder_type="huggingface").

• OpenAI embeddings via the official async client (embedder_type="openai"; install embed-openai and set OPENAI_API_KEY). Use OpenAIEmbedder(..., send_dimensions_to_api=False) when the model has a fixed output size and the API must not get a dimensions field.

• Voyage AI embeddings via aiohttp and the Voyage REST API (embedder_type="voyage"; install embed-voyage and set a Voyage API key). Defaults match VoyageEmbedder in code (voyage-4, 1024 dimensions) when model and dimensions are not set in env.

• Ollama embeddings via the OpenAI-compatible /v1/embeddings endpoint (embedder_type="ollama"; install embed-ollama). Model id and vector dimensions are required in settings so pgvector storage matches the running model.

• PostgreSQL + pgvector for semantic similarity lookup. The library creates a dedicated cache table per embedder configuration (derived from model id and vector dimension) on first use, so you are not tied to a single hard-coded vector width.

• Optional Redis for response caching (keys include an embedder-specific prefix so separate models do not collide). If Redis is not configured, responses are read from Postgres only.

• FastAPI middleware for in-app semantic caching.

• Reverse proxy mode via create_semantic_cache_proxy_app().

Future support

• Django and Flask middleware for in-app semantic caching (not yet shipped; same role as the FastAPI middleware).

Embeddings from the following providers are planned:

• Cohere

Requirements

Python 3.12+.

Links

• Repository: fastapi-semcache

License

Apache-2.0. See LICENSE.