prism-rag 0.2.3

Epistemic Graph RAG with Spreading Activation — retrieval that understands how knowledge relates, not just what it says

PRISM — Epistemic Graph RAG with Spreading Activation

Propagation & Retrieval via Informed Semantic Mapping

PRISM layers a typed epistemic knowledge graph over your existing vector store, then uses spreading activation to surface knowledge structured by how it relates — not just how similar it is.

---

The Problem with Standard RAG

Standard RAG returns a flat ranked list. Every chunk gets a similarity score and nothing else:

query → embed → similarity → [chunk, chunk, chunk, ...]   ← no structure

Chunk B may refute Chunk A. Chunk C may specialise a principle in Chunk D. An older document may have been superseded. Standard RAG can't express any of this.

---

What PRISM Does

PRISM builds a graph where edges carry epistemic type:

Doc A  ──[supports]──▶  Doc B
Doc C  ──[refutes]───▶  Doc D
Doc E  ──[supersedes]▶  Doc F

Retrieval uses spreading activation: a query fires seed nodes via vector search, activation propagates through typed edges, and nodes reached by multiple independent paths (convergence) rank highest.

The result is a structured epistemic answer with five buckets:

Bucket	Contents
PRIMARY	Core relevant chunks, highest convergence
SUPPORTING	Chunks that reinforce or extend the primary answer
CONTRASTING	Chunks that challenge or take a different position
QUALIFYING	Chunks that add conditions, exceptions, or nuances
SUPERSEDED	Historically relevant context now replaced by newer work

---

Installation

pip install prism-rag

Requires Python 3.11+, an existing LanceDB vector store, and an embedding provider.

---

Quick Start

1. Build the epistemic graph (one-time)

from prism import PRISM

# Ollama embeddings (local)
p = PRISM(
    lancedb_path = "/path/to/your/lancedb",
    graph_path   = "/path/to/prism_graph.json.gz",
    ollama_url   = "http://localhost:11434",
    embed_model  = "nomic-embed-text",
    llm_base_url = "https://api.openai.com",
    llm_model    = "gpt-4o-mini",
    llm_api_key  = "sk-...",
)

# Or OpenAI-compatible API embeddings
p = PRISM(
    lancedb_path  = "/path/to/your/lancedb",
    graph_path    = "/path/to/prism_graph.json.gz",
    embed_api_url = "https://api.openai.com/v1/embeddings",
    embed_api_key = "sk-...",
    embed_model   = "text-embedding-3-small",
    llm_base_url  = "https://api.openai.com",
    llm_model     = "gpt-4o-mini",
    llm_api_key   = "sk-...",
)

p.build(k_neighbors=8, cross_source_only=False)

Tip: Use cross_source_only=False (the recommended default). Setting it to True skips intra-document pairs and leaves most epistemic relationships unextracted — on a 30k-chunk corpus this can cut edge count by 3–5×, making the graph too sparse to add value over plain vector search.

Or via the CLI:

prism-build \
    --lancedb-path /path/to/lancedb \
    --graph-path   /path/to/prism_graph.json.gz \
    --llm-api-key  $OPENAI_API_KEY \
    --all-sources

2. Retrieve

p.load_graph()
result = p.retrieve("your question here", top_k=5)
print(result.format_for_llm())

Output:

PRISM retrieval for: "your question here"
────────────────────────────────────────────────────────────

## PRIMARY
[1] source-a  p.14  § 2.1  (score: 0.923)
    The core relevant passage...

## SUPPORTING EVIDENCE
[1] source-c  p.201  § 8.2  (score: 0.841  [via: specializes])
    A passage that extends the primary answer...

## QUALIFICATIONS & NUANCES
[1] source-d  p.38  § 3.1  (score: 0.712  [via: qualifies])
    A passage adding conditions or exceptions...

─ 1 primary · 1 supporting · 0 contrasting · 1 qualifying · 0 superseded ─

3. Access results programmatically

for chunk in result.primary:
    print(chunk.source, chunk.page, chunk.final_score, chunk.text)

for chunk in result.contrasting:
    print("Contrasting view:", chunk.text[:200])

# Feed structured context directly into your LLM
context = result.format_for_llm()

---

Epistemic Edge Types

supports        — A provides evidence reinforcing B
refutes         — A directly contradicts B
supersedes      — A replaces or updates B
derives_from    — A is logically derived from B
specializes     — A is a specific instance of B
contrasts_with  — A and B take different, non-exclusive positions
implements      — A is a concrete method putting B into practice
generalizes     — A is a broader abstraction of which B is a case
exemplifies     — A is a concrete example illustrating B
qualifies       — A adds conditions, exceptions, or nuances to B

Each edge carries a propagation weight (0.40–0.90) and a valence that determines which result bucket its target lands in.

---

Build Performance

The graph is built once offline. PRISM uses a two-stage pipeline that makes large-corpus builds practical:

Stage 1 — Ollama pre-filter (fast, free) An Ollama model screens candidate pairs with a binary yes/no question. ~50% of pairs are discarded before any API call. Runs via Ollama, costs nothing.

PRISM checks the model is available in Ollama before starting and prints a clear warning if not, rather than silently skipping filtering.

Stage 2 — Async LLM classification Surviving pairs are classified with full type + confidence using 20 concurrent API requests, in batches of 20 pairs each.

Build time comparison — 30k-chunk corpus, ~50k candidate pairs:

Pipeline	Wall Time
v1 — sync, batch=5	~40 hours
v2 — async only, batch=20	~30 minutes
v2 — async + stage-1 filter (fast model)	~15–20 minutes

Checkpoint / resume — if interrupted, the build saves progress automatically and resumes from where it left off.

cross_source_only=False produces significantly richer graphs. On a 30k-chunk governance corpus: True = 3,571 edges (graph rarely fires); False = 9,989 edges (supporting/qualifying buckets activate on most queries). Use False unless your sources are genuinely independent.

Choosing a Stage 1 filter model — use a model under ~5 GB. Small, fast models (llama3.1:8b, llama3.2:3b, gemma3:4b) complete each binary call in under a second. Models above ~6 GB — especially over a network connection — can take 2–4 seconds per call and negate the benefit of filtering entirely. If no fast model is available — or if your GPU doesn't have VRAM headroom for true parallel inference — use --no-filter and rely on Stage 2 alone (~30 min).

If your Ollama instance is remote, pass its address via ollama_url:

PRISM(
    ollama_url   = "http://your-ollama-host:11434",
    embed_model  = "qwen3-embedding:4b",
    filter_model = "llama3.1:8b",   # fast model on your Ollama server
    ...
)

Or via CLI:

prism-build --ollama-url http://your-ollama-host:11434 --filter-model llama3.1:8b ...

---

No Re-embedding Required

PRISM works on top of your existing vector store. If you have a LanceDB corpus with embeddings, you don't need to re-index anything.

• Existing vectors → used as-is for seed activation
• Epistemic graph → built from text via LLM, stored as a separate .json.gz file
• Fallback → if no graph exists, PRISM automatically falls back to pure vector search

---

Embedding Providers

Ollama (local):

PRISM(ollama_url="http://localhost:11434", embed_model="nomic-embed-text", ...)

OpenAI-compatible API (OpenAI, Azure, Together, Jina, Mistral, etc.):

PRISM(embed_api_url="https://api.openai.com/v1/embeddings", embed_api_key="sk-...", ...)

---

Links

• Full documentation & architecture: github.com/MadMando/prism
• Issues: github.com/MadMando/prism/issues

---

License

MIT